[0z5a]

InjectPTXAsyncCopy currently strips the base from vector ramps without checking their strides. A 16-byte cp.async then copies contiguous bytes even if the source or destination is a gather/scatter.

Require both ramp strides to be provably one before injecting the instruction, including predicated copies and merged dynamic shared memory. Otherwise keep the original tensor store.

Fixes #20910.

Testing

• On GH200, the original implementation returned [0, 1, 2, 3] for a source ramp selecting [0, 2, 4, 6]; the standalone change produced the expected result.
• The focused test file passed all 7 cases on GH200: 4 IR cases and 3 numerical GPU cases for non-unit source and destination strides. These used the TVM change alone with the pinned TIRx version.
• ruff check, ruff format --check, and clang-format passed for the changed files.