test: stabilize flaky int4-sym fp16 MoE prefill accuracy test

[Copilot]

Description

test_accuracy_int4[False-dtype0] (int4-sym + fp16) in test/test_moe_prefill_accuracy.py fails intermittently on main. The XPU MoE prefill kernel does not guarantee a deterministic reduction order, so long K-reductions (up to K=14336) occasionally produce a few outliers per ~2M outputs that cross the tight fp16 tolerance bound. A pinned RNG seed cannot mask this because the nondeterminism originates in the kernel, not the input.

Changes:

• Widen fp16 tolerance floor — _tol_for_dtype previously applied the 1e-1 floor only to bf16, leaving the known-problematic int4-sym+fp16 path at the tight bound. It now applies the floor to fp16 as well.
• Updated docstring — corrects the stale rationale (fp16 "absorbs" the noise) and documents kernel reduction-order nondeterminism as the actual root cause.

The reference and kernel share the same dequantized weights, so the widened floor still catches real matmul/dequant regressions while absorbing accumulator-order noise. Only quantized paths routed through _tol_for_dtype are affected; the unquantized fp path (_TOL_FP) is unchanged.

Type of Change

Bug fix

Related Issues

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.
• The CUDA CI has passed. You can trigger it by commenting /azp run Unit-Test-CUDA-AutoRound.

[AutoRoundBot]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines could not run because the pipeline triggers exclude this branch/path.