Persistent MXFP8 for experts layers in MoE

[kailash109]

Building off of #1937 and #1898, introduce persistent MXFP8 quantization for expert layers in MoE

Reuses blockwise casting logic from #1898 to write E8M0 scale factors on 1x32 blocks & deferred load of FP8 weights so optimizer states initialized on FP32 master weights, and pack E8M0 scale factors into uint8 bytes for VLLM ModelOpt-MXFP8 offline quantization. _iter_sync_tensors which calls iter_serialized_fp8_tensors now handles both FP8 (hopper-style) and MXFP8 (blackwell-style).

Todo: I still observe perf degradations from mxfp8 vs the bf16 baseline, need to debug why

[erictang000]

qq: how does the reward track vs the bf16 baseline here? also curious about logprobs mismatch

[kailash109]

Reward curves basically identical

and logprob diff seems to be consistently 1.5x that of baseline