[TritonToLinalg](fix) fix zero-stride load boundary

[zhanwei33]

Summary

• Compute a boundary-checked tile size from the original tt.make_tensor_ptr
  logical shape and offsets when any static stride is zero.
• Preserve the existing direct structured lowering (strided<[0, ...]> and
  memref.copy) for those loads; this PR does not add a fallback or change
  later TTAdapter/BishengIR stages.
• Add MLIR coverage for all-zero stride, mixed zero/non-zero stride, negative
  offsets, and the unchanged regular non-zero-stride path.

Root cause

The generic boundary-size helper reconstructs logical coordinates from the
remapped physical pointer. A zero stride removes that physical-coordinate
information, so the reconstruction can reach a 0 / 0 expression and leave a
required OpFoldResult absent. Using the original tensor-pointer shape and
offsets retains the logical coordinates needed for the boundary calculation.

Scope

This is the boundary-only alternative to #1508. It fixes the
triton-to-linalg conversion crash only; it intentionally leaves downstream
acceptance of the resulting zero-stride memref to the existing pipeline. Do
not merge both PRs.

Validation

• pre-commit run --files third_party/ascend/lib/TritonToLinalg/LoadStoreConverter.cpp third_party/ascend/unittest/Conversion/General/TritonToLinalg/zero_stride_block_ptr_load.mlir

• cmake --build <isolated-build> --target TritonToLinalg -- -j 16

• Built a triton-opt driver from this source with the existing TA command-line
  pass registration enabled solely for this check, then ran the new regression:

  triton-opt -allow-unregistered-dialect \
    "--triton-to-linalg=global-kernel=false named-ops=True" \
    --split-input-file \
    third_party/ascend/unittest/Conversion/General/TritonToLinalg/zero_stride_block_ptr_load.mlir \
    | FileCheck \
    third_party/ascend/unittest/Conversion/General/TritonToLinalg/zero_stride_block_ptr_load.mlir