[ROCDL] Gate the TDM tile_dim0 span check on padding

[xudoyuan]

deriveTdmAtom refused any tiled atom whose innermost per-issue span was under 16 bytes. The ISA puts no such floor on tile_dim0. Its inner loop advances by data_size, which is 1, 2, 4 or 8 bytes, and the only rule it states for tile_dim0 is that its span be a whole multiple of 4 bytes -- and only when pad_enable is set, because pad_amount and pad_interval step the LDS pointer in whole DWORDs while bytesStored accumulates across rows, so a row whose span is not a multiple of 4 lands the pad inside the next row. A wider span coalesces better, but the hardware does not require one.

The floor blocked layouts the older make_tdm_atom has always emitted. In aiter, mxfp8's non-preshuffled A-scale tile is (tile_m, tile_k / 32): one e8m0 byte per 32 elements, so 8 bytes at tile_k = 256. That transfer runs correctly through make_tdm_atom today, but make_tiled_tdm_atom could not even construct the op for it.

Checking the 4-byte rule only when geometry.padAmount is set lets that site migrate. Verified with the TDM FileCheck tests (including both negative ones), tests/unit -k "tdm or copy or layout" (86 passed, 1 skipped), and aiter's op_tests/test_gemm_a8w8_mxfp8_bpreshuffle.py with the site switched to the tiled atom: 36 shapes, err 0, over three runs.

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.