[QNN] Enable ConvTranspose + BatchNorm fusion after #23170

[mergennachin]

Feature and motivation

Enable ConvTranspose + BatchNorm fusion in the Qualcomm/QNN pass pipeline after #23170 lands.

FuseBatchNormWithConv.can_fuse currently rejects all transposed convolutions. This workaround avoids the incorrect output-channel scaling in the shared pass discussed in #22994. PR #23170 fixes that shared pass, including grouped transposed weights, but leaves the QNN guard in place. Consequently, QNN will continue to retain standalone BatchNorm operations after that fix.

Proposed work

• Remove the transposed-convolution guard once Fix FuseBatchNormWithConvPass for transposed convolutions #23170 is merged, reusing the corrected shared implementation. Update the subclass documentation or remove the redundant wrapper as appropriate.
• Add Qualcomm regression cases for ConvTranspose + BatchNorm, covering equal and unequal input/output channel counts, grouped convolutions, and convolution bias enabled/disabled where supported by QNN.
• Use nontrivial BatchNorm running statistics and affine parameters. Assert that BatchNorm is folded and the transformed output matches eager execution.
• Run QNN delegation/execution coverage for the supported configurations and retain coverage for ordinary convolution fusion.

Additional context

Local CPU review validation of #23170 with PyTorch 2.13 passed all six new tests; four reproduce the original bug with the pre-fix pass. Eighteen additional grouped/depthwise ConvTranspose1d/2d/3d cases passed through the exported-program transform path. These checks validate the shared pass; QNN integration still needs the coverage above.

cc @cccclai @winskuo-quic @shewu-quic @haowhsu-quic @DannyYuyang-quic @cbilgin @abhinaykukkadapu @psiddh

[shewu-quic]

Hi @mergennachin,
Thanks for raising this issue. We'll take care of it and make sure we have the appropriate coverage in place.