Fix int32 overflow in qdq_mxfp4 causing GPU memory access fault above INT32_MAX elements

[zjin-lcf]

Problem

qdq_mxfp4 / qdq_mxfp4_ truncate numel() into an int and index elements with an int. On any tensor with more than INT32_MAX (2,147,483,647) elements the count wraps negative, still passes the divisibility check, and the kernel launches with a garbage grid dimension. The process dies with:

[zjin-lcf]

Updated per review.

1. Is numel a multiple of block_size?

Yes — guaranteed by the precondition immediately above the grid computation, which selects block_size from {128, 64} precisely so the division is exact and rejects anything else:

2. Grid-stride loop

Agreed, and switched to it. The previous revision converted one failure mode into another — it raised an error when grid_size > INT_MAX instead of removing the ceiling. qdq_mxfp4_kernel is now a grid-stride loop and the launcher clamps the grid, so the limit is gone rather than relocated. For any tensor that fits in memory today the clamp is inactive and each thread still handles exactly one element, so there is no change to the common path.

Verification (MI300A / gfx942, ROCm 7.2), differential against the stock kernel rather than a hand-written reference:

test	result
8,388,608 elements, mixed magnitudes (1e-3 to 1e3)	bit-identical to stock
same input, grid clamped to 1024 blocks → 64 grid-stride iterations per thread	bit-identical to stock
2,132,795,392 elements (under INT32_MAX)	OK before and after
2,162,155,520 elements (over INT32_MAX)	faulted before, passes now