[aseemw]

This PR fixes two cases where graph-mode quantization silently left conv/linear weights unquantized.

Issue 1: convs with padding="same" or padding="valid" were not quantized

Cause: torch.export exports a conv with string padding as a different aten overload, torch.ops.aten.conv2d.padding instead of torch.ops.aten.conv2d.default (likewise for conv1d/conv3d). The conv annotation patterns only contained aten.convNd.default, so they never matched these convs.

Fix: the conv patterns now also include the aten.convNd.padding overload.

Issue 2: conv/linear followed by an in-place activation (e.g. nn.ReLU(inplace=True)) was not quantized

Cause: match_pattern_with_subgraph_matcher keyed each match on the "mod" entry of the pattern's name node map, expected to point to the weighted op (the conv/linear). With In-place activation present, torch.export maps "mod" to the relu_ node instead of the conv. The match was then keyed on relu_, which received the input and weight specs it has no use for, and the conv got neither its input nor its weight quantizer.

Fix: the primary node of a match is now the op in the match that consumes the pattern's "weight", rather than the node named "mod".

[NehalBhandari]

Thank you for the fixes. LGTM!

One thing to note is that we are not handling conv with padding in _conv_bn_utils which handles folding of conv bn weights. But torchao doesn't handle padding ops in fusion either so it should be okay to keep batch norm separate.