[ROCm][CI] Stabilize MI355 FlyDSL MoE accuracy test

[AndreasKaratzas]

This PR stabilizes the two MI355 FlyDSL W4A16 MoE cases that failed in AMD CI build 12250. The test was introduced in #44400, and its test, kernel, and conversion code are unchanged between recent passing and failing builds; #41100 changed test collection and exposed the test's dependence on inherited RNG state. The failures contained only a few BF16 accumulation outliers among as many as 117 million output values, so a strict torch.allclose made the largest cases brittle. This change seeds the inputs and checks both the mismatch rate and worst normalized error, retaining sensitivity to broad or severe accuracy regressions.

• Fix: Seed every parameterized case, require at least 99.999% of values to satisfy the original atol=0.5, rtol=0.1 tolerance, and bound every permitted outlier to 3x its elementwise tolerance.
• Repeated validation: 30/30 independent full-file runs passed across two gfx950 MI355 GPUs, for 900/900 collected test cases.
• Input sweep: 30 distinct seeds passed both formerly failing shapes in fresh processes across two additional MI355 GPUs, for 60/60 exact-shape runs.
• Control: With the original unseeded strict assertion restored, the two cases failed in 3/10 fresh processes; deterministic diagnostics observed only 0-3 mismatches out of 117,440,512 values, with a worst error of 1.086x the original elementwise tolerance.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[AndreasKaratzas]

/ci run

[github-actions]

✅ Triggered Buildkite CI #84715 for commit b7a77e2ea2de.