feat(sglang): NVFP4 quantized rollout and refit

[xiuhu17]

Adds scheme: nvfp4 alongside the existing mxfp8 path under
policy.generation.sglang_cfg.quantization, so SGLang rollouts can run
NVFP4 weights while training stays BF16/FP8.

Stacked on zhw/mxfp8_support; review only the top commit.

What this does

• nvfp4_quantization_core.py — TE-backed NVFP4 weight quantization.
  Two-level scaling (per-16-element E4M3 block scales + a per-tensor global
  encode/decode scale), row padding to TE's 16-row alignment, and gate/up
  pair quantization so a fused gate_up_proj gets one shared global scale
  instead of two divergent ones.
• nvfp4_setup.py — offline HF→NVFP4 checkpoint conversion with cache
  fingerprinting, so SGLang boots from a checkpoint whose tensor layout and
  high-precision exclusions match what the online refit will send.
• quantization_utils.py — the scheme-agnostic pieces both schemes need:
  scheme validation, the dynamic high-precision policy
  (extra_high_precision_layers_hf, modules_to_not_convert,
  num_layers_at_start_in_bf16 / ..._end_in_bf16), and fused-MoE container
  expansion. SGLang picks one quantization method per FusedMoE module, so a
  single skipped expert tensor has to pull its whole container into BF16.
• Refit session API — post_process_weights() is replaced by
  begin_weight_update() / end_weight_update(). Quantized weights carry
  derived scales that have to be re-materialized after the last bucket
  lands, which the old single-shot post-process could not express.
  utils/patches.py backports the endpoints to the pinned SGLang release.
• MegatronSGLangHfWeightIterator — quantizes during the AutoBridge walk
  and keeps every companion tensor produced from one source weight (and each
  NVFP4 gate/up pair) inside the same bucket, so a bucket boundary can never
  split a weight from its scales.
• TransformerEngine is bumped to v2.17 for the NVFP4 quantizer API.

The DTensor/FSDP policy stays BF16-only and now says so explicitly
(dtensor_sglang_refit.refit_sglang_colocated raises on any other scheme)
rather than silently sending BF16 to a quantized engine.

Prior art

This follows the NVFP4 RL work in Miles:

• [Roadmap] Blackwell MXFP8 and NVFP4 RL training radixark/miles#615 — the tracking issue
• https://humansand.ai/blog/nvfp4-rl — the writeup
• NVFP4 RL radixark/miles#1261 — the Miles implementation, by @zianglih

Status

Draft. Unit tests for the quantization core, the offline converters, the
skip/high-precision policy, the weight iterator, and the SGLang patches are
included and pass. uv.lock was hand-resolved for the TE bump during the
rebase onto main and should be regenerated with uv lock in the project
container before this merges.

[xiuhu17]

Opened against the wrong repo — reopened upstream as NVIDIA-NeMo#3426.