fix(sm90): sanitize pruned FP8 MegaMoE weight blocks

[ercaoeatting]

Summary

SM90 MegaMoE can produce incorrect output when a pruned 128x128 FP8 weight block retains arbitrary FP8 bytes while its weight scale is tiny but nonzero.

This change sanitizes those blocks during transform_weights_for_mega_moe_sm90:

• Treat scales with abs(scale) < 1e-12 as pruned.
• Zero the corresponding FP8 weight block before the SM90 fused kernel sees it.
• Replace the tiny scale with the expert's median normal scale, or a finite fallback when the expert is entirely pruned.
• Clone only when sanitization is needed, so caller-owned tensors are not modified and the existing no-tiny-scale path keeps its identity/behavior.

The same handling is applied to both W13 and W2 weights.

Root cause

The SM90 MegaMoE split-MN path rescales partial results with reciprocals of the weight scales. A tiny nonzero scale can therefore amplify residual FP8 bytes in a block into a large incorrect result. A truly pruned block must be represented as zero data with a safe finite scale.

Validation

On an H20D (SM90) in the container:

• test_mega_moe_sm90_weight_transform.py --num-processes 1 — PASS
• test_mega_moe_pre_dispatch_sm90.py --num-processes 1 — 8 cases PASS
• SM90 fused-only MegaMoE smoke (tokens=1, experts=1, topk=1) — kernel launched successfully (52.5 us, non-fatal symmetric-memory multicast warning)
• The prior minimal repro produced max_abs=1870659584.0 with the unsanitized tiny-scale block and max_abs=0.0 after sanitization.

The SGLang-side workaround can be removed once this fix is available in a released DeepGEMM package.