convert _upsample_bilinear2d_aa with the antialias filter - #2907

_upsample_bilinear2d_aa was registered as an alias of upsample_bilinear2d , so antialias=True was dropped without a warning . torchvision's Resize uses it by default , so most image models that resize on the way in get it . in torch the two only agree on an upsample without align_corners , every downsample and everything with align_corners=True is different .

this sends the _aa kinds through the separable path from #2818 , with aten's triangle filter in place of the cubic one . the weight code is shared , only the filter and the tap count change .

that was not enough on its own . at an image size the separable path itself comes out wrong on the core ml runtime , for bicubic on main as well . a 512x384 to 224x168 resize was off by up to 0.87 for bicubic , while the same weights applied in numpy match torch to 1e-5 . i narrowed it to the gather per tap , then mul and reduce_sum over the taps . mul alone and reduce_sum alone are exact , the two together are not once the tensor gets past about 65k elements with a short last axis , filed as #2906 :

@mb.program(input_specs=[mb.TensorSpec(shape=(98, 168, 4)), mb.TensorSpec(shape=(98, 168, 4))], opset_version=ct.target.macOS15)
def prog(x, y):
    return mb.reduce_sum(x=mb.mul(x=x, y=y), axes=[-1])

(96, 168, 4) is exact and (98, 168, 4) is off by about 3 , cpu only , fp32 , empty pass pipeline . so the third commit scatters the taps into a dense (out, in) matrix and applies it with one matmul per axis , which is also fewer ops .

torchvision.transforms.v2.Resize((224, 168)) on a 512x384 image in [0, 1] was off by 0.116 on average and up to 0.497 on main . with this it is within 7.3e-06 of torch .

a dynamic shape upsample without align_corners is the same with or without antialias , so it still goes to the plain bilinear lowering , which handles dynamic shapes . a dynamic downsample raises now instead of converting to the wrong answer .

44 of the new antialias cases fail without the converter change , and the image size test fails 4 of 9 with the gather version , plain bicubic included . TestUpsample with INCLUDE_MIL_FP32_UNIT_TESTS=1 is 945 passed , 3 skipped , 99 xfailed .