[Torch] Canonicalize aten.convolution with single int tuple params

[hariprasadravi]

This PR is a more robust fix for the issue captured in #4380 .

Essentially, torch.ops.aten.convolution lowering to tosa, linalg and stablehlo fail if stride, padding, dilation or output_padding is a tuple with a singleton element, but convolution operates on 2 or 3 spatial dimensions.

In the failing case, torch.nn.Conv2d with padding= 'valid' generates a torch.ops.aten.conv2d.padding op in ExportedProgram, which is later decomposed to torch.ops.aten.convolution.default with a single padding value of [0] after running ep.run_decompositions().

In #4380, I attempted to fix just the torch-to-tosa pass, but I later realised that this is a more general bug in all the backends for multiple params (thanks to #4380 (comment)).

Fix:

I followed #4250 to canonicalize aten.convolution if it operates on 2 or 3 spatial dims but params are singleton. For example, if aten.convolution is 2D but padding == [0], we canonicalize it to padding == [0, 0]

[hariprasadravi]

@sahas3 @sjarus @zjgarvey and @vivekkhandelwal1 . Please let me know if you have any thoughts.