[Torch] Add aten.xlogy.Tensor decomposition

[josephbak]

aten.xlogy.Tensor had no lowering on any backend. It also blocked the torchdynamo config of the KlDivLoss e2e tests, since PyTorch's kl_div decomposition emits xlogy (see #4204, which added kl_div and recorded the gap).

This decomposes xlogy into log, mul, eq.Scalar, isnan and where, so all backend configs are handled at once. Shape and dtype functions are added as well; integer inputs promote to float32 since log forces a float.

This builds on the abandoned #2486 with its review feedback applied. I also fixed the type derivation, which was wrong under broadcasting: eq.Scalar, log and isnan are shape-preserving, so their result types have to come from their operands, not the broadcast result. This is the same issue that #4509 fixed for logaddexp. A [3,1] x [1,5] e2e test covers it.

The broadcast test uses static shapes because TorchToLinalg's elementwise lowering only emits broadcast maps when size-1 dims are statically known; with dynamic sizes it emits equality assertions instead. This is pre-existing and general (add.Tensor / mul.Tensor behave the same way).

Testing

• Lit test for the decomposition.
• e2e tests pass on linalg, fx_importer, fx_importer_stablehlo, and fx_importer_tosa (built StableHLO locally to verify that config).
• Xfailed on the onnx config: PyTorch's ONNX exporter doesn't support aten::xlogy (UnsupportedOperatorError).
• The 5 KlDivLoss tests are removed from TORCHDYNAMO_CRASHING_SET; they were skipped entirely before and now run and pass under --config=torchdynamo.

[josephbak]

Hi @zjgarvey — could you (or another maintainer) approve the CI workflows here? This picks up the abandoned #2486 and unblocks the KlDivLoss torchdynamo tests recorded in TORCHDYNAMO_CRASHING_SET by #4204, which you reviewed and merged. Happy to take a review from whoever has time.

[josephbak]

Friendly ping — this has been green for several days (checks passing, no conflicts). @zjgarvey @sahas3 — would either of you have a moment to review, or point me to whoever's handling torch decompositions these days? This builds on #4204 and unblocks its KlDivLoss torchdynamo tests.

Happy to address changes. Thanks!

[josephbak]

@sahas3 the nightly run failed at "Setup workspace and python" with a transient "Service Unavailable" fetching an action, before any build or test ran (stable was auto-cancelled). Looks like CI infra, not the change. Could you re-run when you get a chance? Thanks!

[sahas3]

LGTM.

[josephbak]

Thanks for the review and guidance on the approach, @sahas3!