What does this PR do?

Closes #3150 by:

• Using Welford's algorithm for _pearson_corrcoef_update
• Tracking maximum absolute deviation (approximately) and uses them for scaling in _pearson_corrcoef_compute

from torchmetrics.regression import PearsonCorrCoef
from torchmetrics.functional.regression.pearson import pearson_corrcoef
import torch
from scipy.stats import pearsonr

torch.manual_seed(42)
scale = 1e-4
bias = 1e2
preds = torch.randn(100)
target = preds + 0.1 * torch.randn(100)
preds *= scale
target *= scale
preds += bias
target += bias
print(pearson_corrcoef(preds, target))
print(pearsonr(preds.cpu().numpy(), target.cpu().numpy())[0])

For this test, where predictions and targets are large in their absolute values but have small variances,

Original implementation: nan
Current implementation: 0.9955
Scipy reference: 0.99274683 

For above test with scale = 1e-3 and bias = 1e1,

Original implementation: nan
Current implementation: 0.9960
Scipy reference: 0.99597996

It is not as accurate as scipy, but for a streaming algorithm I think this is reasonably good.

Especially when bias = 0, current implementation using torch.float32 works up to scale=1e-18. (original implementation fails at scale = 1e-4)

Note: The tests were modified for current implementation. Specifically, inputs in test_pearsons_warning_on_small_input are changed, and for test_final_aggregation_no_inplace_change, the test logic incorporates max_abs_dev_*.

Did you have fun?

Make sure you had fun coding 🙃

📚 Documentation preview 📚: https://torchmetrics--3152.org.readthedocs.build/en/3152/