[torchlib] Support divisor_override in avg_pool2d and avg_pool3d

[Raashish Aggarwal (raashish1601)]

aten_avg_pool2d and aten_avg_pool3d accepted divisor_override but ignored it (there was a TODO for it), so the exported model silently divided by the window size instead:

x = torch.arange(16.0).reshape(1, 1, 4, 4)
F.avg_pool2d(x, 2, 2, divisor_override=1)  # [10, 18, 42, 50]
# exported model returns [2.5, 4.5, 10.5, 12.5]

AveragePool can't take a custom divisor, so with divisor_override the window sums are computed with LpPool(p=1) on the positive and negative parts (sum = LpPool(relu(x)) - LpPool(relu(-x))) and divided by divisor_override. Padding contributes zeros and LpPool supports ceil_mode, so this matches PyTorch, which divides the sum of the in-bounds values by divisor_override regardless of count_include_pad. Without divisor_override the graph is unchanged.

Testing:

• Removed the divisor_override xfails for nn.functional.avg_pool2d / avg_pool3d, and excluded divisor_override samples from the avg_pool3d ceil_mode xfail, since they no longer go through AveragePool. pytest tests/function_libs/torch_lib/ops_test.py -k avg_pool passes with onnxruntime 1.23.0 (the version in noxfile.py): 5 passed, 33 subtests passed, 7 xfailed.
• I also compared LpPool-based sums with F.avg_pool2d(..., divisor_override=3) over 96 combinations of input size, kernel, stride, padding, ceil_mode and count_include_pad on onnxruntime 1.23.0 and 1.30.0. All match. Exporting nn.AvgPool2d(..., divisor_override=...) and 3d/unbatched cases with torch.onnx.export(dynamo=True) also matches eager.
• ruff check and format are clean on the changed files.