That significantly improves performance and addresses correctness problem(to an extend permitted by reducing precision of scale factor computation to float32). uint8 scaling algorithm mimics CPU/Pillow implementation
https://github.com/python-pillow/Pillow/blob/569b785371aa717a004adb0166feb565bbb01b7b/src/libImaging/Resample.c#L306-L309
I.e. using fixed precision integral arithmetic and rounding results of horizontal interpolation back to integers before performing vertical one, which results in technically less accurate results.

But even with those changes, `atol`, `rtol` must be tweaked to `1, 0` when scale factor is `1/3` or `2/3` because of the difference of representation  of those values as floats and doubles.

Changes in the performance could be measured using the following script
```python
import torch
import time
import subprocess

def benchmark(device, dtype):
    # Create example inputs
    x = torch.testing.make_tensor(1, 1, 2048, 2048, device=device, dtype=dtype)
    sf = .5

    # Check output
    y = torch.nn.functional.interpolate(x, scale_factor=sf, mode="bilinear")
    z = torch.nn.functional.interpolate(x.cpu(), scale_factor=sf, mode="bilinear")
    outputs_match = torch.allclose(y.cpu(), z)
    if not outputs_match:
       atol = (y.cpu() - z).abs().max()
       rtol = ((y.cpu() - z)[z!=0]/z[z!=0]).abs().max()
       print(f"atol={atol} rtol={rtol}")

    # Measure time manually
    start_time = time.time() * 1000
    for _ in range(1000):
        y = torch.nn.functional.interpolate(x, scale_factor=sf, mode="bilinear")
    torch.mps.synchronize
    end_time = time.time() * 1000
    manual_delta = (end_time - start_time)
    average_time = f"{manual_delta:6.1f}"

    return "True " if outputs_match else "False", average_time

outputs_match_list = []
average_time_list = []
for device in ["mps", "cpu"]:
    for dtype in [torch.float32, torch.float16, torch.bfloat16, torch.uint8]:
        outputs_match, average_time = benchmark(device, dtype)
        outputs_match_list.append(str(outputs_match))
        average_time_list.append(average_time)

brand_string = subprocess.check_output(['sysctl', '-n', 'machdep.cpu.brand_string']).decode("utf-8").strip()
print(f"\nBenchmarking Results (collected on {brand_string}):")
print("-"*40)
print("Device            :                MPS                 |               CPU")
print("Dtype             :   FP32  |  FP16  |  BF16  |   U8   |  FP32  |  FP16  |  BF16  |  U8")
print(f"Outputs Match     :  ", " |  ".join(outputs_match_list))
print(f"Average Time (us) :", "  |".join(average_time_list))
```

Benchmark results before
```
Benchmarking Results (collected on Apple M4 Pro):
----------------------------------------
Device            :                MPS                 |               CPU
Dtype             :   FP32  |  FP16  |  BF16  |   U8   |  FP32  |  FP16  |  BF16  |  U8
Outputs Match     :   True  |  True  |  True  |  False |  True  |  True  |  True  |  True
Average Time (us) :  277.3  | 197.2  | 188.0  | 163.5  | 302.8  | 248.1  | 308.7  | 650.9
```
After(almost **100x** perf gain):
```
Benchmarking Results (collected on Apple M4 Pro):
----------------------------------------
Device            :                MPS                 |               CPU
Dtype             :   FP32  |  FP16  |  BF16  |   U8   |  FP32  |  FP16  |  BF16  |  U8
Outputs Match     :   True  |  True  |  True  |  True  |  True  |  True  |  True  |  True
Average Time (us) :    1.7  |   1.5  |   1.7  |   1.5  | 296.5  | 236.0  | 310.8  | 642.6
```
Pull Request resolved: #145581
Approved by: https://github.com/Skylion007
ghstack dependencies: #145578

That significantly improves performance and addresses correctness problem(to an extend permitted by reducing precision of scale factor computation to float32). uint8 scaling algorithm mimics CPU/Pillow implementation
https://github.com/python-pillow/Pillow/blob/569b785371aa717a004adb0166feb565bbb01b7b/src/libImaging/Resample.c#L306-L309
I.e. using fixed precision integral arithmetic and rounding results of horizontal interpolation back to integers before performing vertical one, which results in technically less accurate results.

But even with those changes, `atol`, `rtol` must be tweaked to `1, 0` when scale factor is `1/3` or `2/3` because of the difference of representation  of those values as floats and doubles.

Changes in the performance could be measured using the following script
```python
import torch
import time
import subprocess

def benchmark(device, dtype):
    # Create example inputs
    x = torch.testing.make_tensor(1, 1, 2048, 2048, device=device, dtype=dtype)
    sf = .5

    # Check output
    y = torch.nn.functional.interpolate(x, scale_factor=sf, mode="bilinear")
    z = torch.nn.functional.interpolate(x.cpu(), scale_factor=sf, mode="bilinear")
    outputs_match = torch.allclose(y.cpu(), z)
    if not outputs_match:
       atol = (y.cpu() - z).abs().max()
       rtol = ((y.cpu() - z)[z!=0]/z[z!=0]).abs().max()
       print(f"atol={atol} rtol={rtol}")

    # Measure time manually
    start_time = time.time() * 1000
    for _ in range(1000):
        y = torch.nn.functional.interpolate(x, scale_factor=sf, mode="bilinear")
    torch.mps.synchronize
    end_time = time.time() * 1000
    manual_delta = (end_time - start_time)
    average_time = f"{manual_delta:6.1f}"

    return "True " if outputs_match else "False", average_time

outputs_match_list = []
average_time_list = []
for device in ["mps", "cpu"]:
    for dtype in [torch.float32, torch.float16, torch.bfloat16, torch.uint8]:
        outputs_match, average_time = benchmark(device, dtype)
        outputs_match_list.append(str(outputs_match))
        average_time_list.append(average_time)

brand_string = subprocess.check_output(['sysctl', '-n', 'machdep.cpu.brand_string']).decode("utf-8").strip()
print(f"\nBenchmarking Results (collected on {brand_string}):")
print("-"*40)
print("Device            :                MPS                 |               CPU")
print("Dtype             :   FP32  |  FP16  |  BF16  |   U8   |  FP32  |  FP16  |  BF16  |  U8")
print(f"Outputs Match     :  ", " |  ".join(outputs_match_list))
print(f"Average Time (us) :", "  |".join(average_time_list))
```

Benchmark results before
```
Benchmarking Results (collected on Apple M4 Pro):
----------------------------------------
Device            :                MPS                 |               CPU
Dtype             :   FP32  |  FP16  |  BF16  |   U8   |  FP32  |  FP16  |  BF16  |  U8
Outputs Match     :   True  |  True  |  True  |  False |  True  |  True  |  True  |  True
Average Time (us) :  277.3  | 197.2  | 188.0  | 163.5  | 302.8  | 248.1  | 308.7  | 650.9
```
After(almost **100x** perf gain):
```
Benchmarking Results (collected on Apple M4 Pro):
----------------------------------------
Device            :                MPS                 |               CPU
Dtype             :   FP32  |  FP16  |  BF16  |   U8   |  FP32  |  FP16  |  BF16  |  U8
Outputs Match     :   True  |  True  |  True  |  True  |  True  |  True  |  True  |  True
Average Time (us) :    1.7  |   1.5  |   1.7  |   1.5  | 296.5  | 236.0  | 310.8  | 642.6
```
Pull Request resolved: pytorch#145581
Approved by: https://github.com/Skylion007
ghstack dependencies: pytorch#145578