[WIP] support value head

[PengchengShi00]

Summary

• Add a critic ValueHead (nn.Linear(hidden, 1)) selected by TransformerConfig.head_type="value_head". Runtime module name stays lm_head; HF save/load uses value_head.weight. Compose/VL only needs text_config.head_type.
• Add CriticLossConfig: SFT MSE when old_values is omitted, OpenAI-style clipped value loss when it is provided. value_head rejects MTP and does not tie embeddings.
• Rename FSDPConfig.fp32_lm_head → fp32_head (covers LM and value heads). RolloutConfig.fp32_lm_head is unchanged. Actor checkpoints load into a critic with strict=False; missing value-head weights are randomly initialized on all ranks.

Out of scope

• Trainer default lm_loss_cfg is still CELossConfig; critic is not wired into TrainEngine / Trainer._prepare_model_input.
• TrainingWorker.fit remains the actor (advantages / logprobs) path. WorkerConfig.loss_cfg accepts CriticLossConfig for typing only.
• No CriticWorker, GAE, or vision-side value head.

Test plan

• tests/engine/test_critic_sft.py with QWEN3_5_MOE_PATH (Qwen3.5-35B-A3B): MTP reject, HF key remap, SFT MSE, clipped value loss, actor→critic load / save_hf / reload
• Existing RL configs still construct FSDPConfig(fp32_head=...)