test(moe): grant one bf16 ulp only at rounding ties in the ds_fp4 transcription check

[void-mckenzie]

Closes #649.

The transcription check rounds each route's output to bf16 before the fp32 sum. On
sm_120 the kernel's accumulation order puts token 1 / route 2 / column 493 of the
seeded inputs at -97.75000x, 1e-5 past the bf16 tie between -97.5 and -98.0, while
the torch reference lands 1e-5 short of it: the same value rounds to -98.0 on one
side and -97.5 on the other. The four routes cancel to -4.5, so the one-ulp (0.5)
difference exceeds the allowed 0.11 (atol + rtol*|sum| with the flat 2e-2).

reference() now also returns a per-element tie budget: the largest bf16 ulp among
routes whose raw output sits within 0.1 ulp of a rounding tie, which is the only
place a ~1e-5 relative difference in the fp32 accumulation can change the rounded
result. _assert_close widens atol + rtol*|want| by that budget, so every element
away from a tie keeps exactly the 2e-2 check. The grouped-prefill, cpu-executor and
inactive-routes tests compare the same sum and go through the same helper.

Tested on RTX 5090 (sm_120), driver 580.95.05, torch 2.11.0+cu130, python 3.10.19;
synthetic banks, no checkpoint:

.venv/bin/python -m pytest tests/moe/test_dsfp4_reference_semantics.py -q

• main 95ac86f: test_gemv_path_matches_the_transcription[32] and [128] fail,
  1 / 3072 elements, abs 0.5 at (1, 493) - identical to the issue
• this branch: 7 passed