[DO NOT REVIEW] Add MXFP8 recipes for the qwen3 and gpt_oss MoE models

[anijain2305]

The grouped-expert support landed without a way to turn it on outside
DeepSeek-V3, so the only model exercising it was the one it was developed
against. Add debug recipes for the two other MoE models whose experts the
converter already handles.

Both quantize the attention projections and the expert grouped GEMMs, and
leave the router gate and lm_head in BF16: the gate is tiny and its output
drives routing, and the vocab projection is large and numerically sensitive.
Only the fused QKV saves a quantized input activation for WGRAD, since
nothing else consumes it, while attention already retains the output
projection's input.

gpt_oss exercises a path DeepSeek-V3 does not: its experts carry per-expert
biases and a SwiGLU clamp, and the converter preserves both by subclassing
GptOssGroupedExperts rather than replacing it -- only the grouped GEMM seam
changes, and the 2D biases are skipped when installing compute weights.

Add CPU coverage for both recipes. MXFP8 needs SM100, which CI does not have,
but the config-tree transform does not, so the test bypasses the hardware
gate and asserts which modules are converted, which input-activation format
each linear saves, and that the experts subclass the model's own variant.

Verified on 2xGB200: qwen3_moe_debug_mxfp8 trains with loss tracking BF16
(6.673 vs 6.603 at step 3) and 0.6 GiB lower peak memory.

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}