[nxphi47]

Add native MXFP8 QAT for dense/grouped projections and routed experts. Forward uses DeepGEMM per-32 MXFP8 GEMMs; backward uses BF16 GEMMs with the forward-rounded weights for input gradients and original BF16 activations for weight gradients.

Add deterministic Triton stochastic weight rounding keyed by step, tensor identity, and global expert offset. Reuse the same weight sample for backward and activation-checkpoint replay. Activations remain nearest-rounded. The vLLM compatibility patch uses that same per-32 cast only in online dense/expert weight loading, including checkpoint reload; serving draws fresh streams independently from training. Add explicit configuration and reject incompatible blockwise-FP8 options for MXFP8.

Stacked on #4092. The kernel submodule pins PrimeIntellect-ai/prime-kernels#19; the release wheel does not yet include MXFP8, so the pinned kernel source must be installed using the repository's source-build workflow.

Validation completed:

• 14 kernel tests, 25 MXFP8 trainer tests, 17 existing FP8-rounding regression tests, and 2 inference-cast tests passed.
• Three distributed tests passed on two GPUs, covering FSDP, split-token inputs, and expert shards.
• Real vLLM dense/MoE model loading, generation, and layerwise checkpoint reload passed for RN and SR; 22 weight casts were checked on each initial load and again on each reload.
• Lint and whitespace checks passed.

Draft pending full-model integration validation. Optimizer and reduction dtypes are unchanged.