Add NVFP4 quantization-aware distillation (QAD)

[andrewor14]

Summary: QAD recovers the accuracy lost when a model is quantized to NVFP4 by distilling from the original bf16 model into an NVFP4 fake-quantized copy. The student's MoE experts and dense linear layers are fake-quantized with a straight-through estimator during training; a frozen bf16 teacher supervises via KL divergence (plus an optional hard-label cross-entropy term).

Adds:

• NVFP4 fake-quant primitives: (can be moved to torchao if needed) torchtitan/components/quantization/ nvfp4_fake_quant.py -- self-contained two-level block-scale E2M1 fake-quant (STE), plus a single entry point apply_nvfp4_fake_quant(model) that quantizes both the MoE experts and the dense linears (excluding lm_head/router/gate).
• KD/QAD trainer + configs: torchtitan/experiments/kd/ -- KDTrainer (frozen bf16 teacher + fake-quant student), kd_loss (KL + CE), and qwen3 QAD configs (debug + 30B-A3B).
• MoE fake-quant hook: GroupedExperts.fake_quant_fn in models/common/moe.py, applied to expert weights and activations inside _experts_forward.
• Post-init hook: post_model_init_fn in trainer.py, used to apply the fake-quant to the student after model init/parallelization.
• Single-file HF checkpoint support: protocols/state_dict_adapter.py.

Test Plan:

MODULE=kd.qwen3 CONFIG=qwen3_30b_a3b_qad NGPU=8 ./run_train.sh \
    --training.steps 200 \
    --checkpoint.interval 50 \
    --checkpoint.initial_load_path /path/to/qwen3-30b-a3b-hf \
    --hf_assets_path /path/to/qwen3-30b-a3b-hf

Quick self-contained smoke test (tiny debug model, no external checkpoint):

MODULE=kd.qwen3 CONFIG=qwen3_moe_debug_qad NGPU=2 ./run_train.sh

Results (MATH accuracy, bf16 -> nvfp4):

  qwen3-30B-A3B
    baseline (200-step RL ckpt):  88.6% -> 86.6%   (-2.0)
    + QAD (800 steps):            nvfp4 88.6%       (+2.0)
    -> 100% of the gap recovered

  gpt-oss-20B
    baseline (100-step RL ckpt):  62.6% -> 50.8%   (-11.8)
    + QAD (1400 steps):           nvfp4 55.2%       (+4.4)
    -> 37% of the gap recovered

QAD closes the full NVFP4 gap on qwen3-30B-A3B and recovers a third of the (much larger) gap on gpt-oss-20B.

[pytorch-bot]

Workflows were awaiting approval. CI has now been triggered for the ciflow labels on this PR.