Fix SIGABRT on zero-extent slice copy (SliceCopyStartGreaterThanDimSize)

[alexxony]

Summary

Fixes the SIGABRT crash in SliceCopyStartGreaterThanDimSize_Module_basic
(linalg-on-tensors e2e test). The crash happens because x[100:10:1] = y
with y = torch.rand(0, 4, 4) (an empty slice source) hits two independent
bugs in the lowering/runtime path. Both are fixed here; each is a small,
localized commit.

Commit 1: TorchToTMTensor broadcast fold (numpy semantics bug)

getBroadcastShape (lib/Conversion/TorchToTMTensor/TorchToTMTensor.cpp)
folds index-tensor broadcast sizes with max(size, acc), where acc
starts at 1. This maps a size-0 dimension (from an empty index/slice) to
1 instead of preserving 0 — the numpy broadcasting rule is
(size == 1) ? acc : size, which only differs from max at 0. This
corrupted the scatter loop bound for index_put lowering derived from
RecomposeSliceCopy_.

Commit 2: RefBackend zero-extent memref stride canonicalization

Separately, torch.rand(0, 4, 4).numpy() reports strides (0, 0, 0).
When RefBackend marshals this into a memref descriptor as-is, the
runtime-verification pass's memref.cast check (identity/contiguous
layout, last-dim stride == 1) rejects it and aborts — even though the
array has zero addressable elements, so the stride is semantically
irrelevant.

This is the same fix as #4680, which was self-closed by the author without
review on 2026-07-31. Looking at that author's other PRs closed the same
day, the closures were driven by a TOSA-specific realization (zero-extent
tensors aren't TOSA-conformant) — #4680 targets RefBackend
(linalg-on-tensors), which has no such constraint, so it appears to have
been swept up in that batch close rather than rejected on its own merits.
I independently hit the same bug, arrived at the same fix, and am
reviving it here with a regression test.

Commit 3: keep the ONNX xfail set consistent

ONNX_CRASHING_SET inherits LINALG_CRASHING_SET, so removing the four
zero-extent tests from the latter makes the onnx config (which CI always
runs) attempt them. Running it showed:

• TraceModule_empty and TraceUnsignedIntModule_empty now pass, so they
  are removed from ONNX_XFAIL_SET (they would otherwise be unexpected
  passes).
• SliceCopyStartGreaterThanDimSize_Module_basic fails under onnx in
  torch.onnx.export (inconsistent constraints for the empty slice), so it
  is added to ONNX_XFAIL_SET.
• ReduceAllDimEmpty_basic passes under onnx and needs no change.

Testing

• New lit test: projects/pt1/python/test/refbackend/zero_extent_memref.py
• check-torch_mlir-conversion-torchtotmtensor: 1/1
• check-torch_mlir-conversion-torchtolinalg: 22/22 (no regressions)
• SliceCopyStartGreaterThanDimSize_Module_basic via
  projects.pt1.e2e_testing.main -c linalg: now PASSES (was SIGABRT)
• Full, unfiltered runs of the four configs CI uses (onnx, fx_importer,
  fx_importer_stablehlo, fx_importer_tosa) on this branch: no crashes.
  The only non-passing results are 22 unexpected passes under onnx and 3
  failures under fx_importer_tosa (ArgminIntModule_basic,
  ArgminIntModule_multiple_mins, ReduceMinAlongDimSignedInt_basic), and
  the same ones occur on the base commit, so they are not caused by this PR.
  The four tests above end up PASS/XFAIL as intended in each config (torch
  2.15.0.dev20260917).

Scope note

SliceCopyStartGreaterThanDimSize_Module_basic also appears in xfail
lists for other backends (e.g. TOSA/STABLEHLO sections) — those are
not touched here. The stablehlo and tosa sets do not inherit
LINALG_CRASHING_SET, and with this change the tests stay expected-fail
there. This PR removes the entries from the linalg-on-tensors
LINALG_CRASHING_SET, where they were directly verified, plus the ONNX
adjustment above.

Assisted-by: Claude (Anthropic)

[alexxony]

Friendly note: CI shows action_required (needs a maintainer to approve the workflow run for first-time contributors) rather than having actually run. Happy to address any review feedback once CI is green.