feat(nano): support LoRA finetuning of the Qwen3 LLM

[KamitobiHaru]

Summary

The FunASR-Nano model config already ships llm_conf.use_lora / llm_conf.lora_conf
(r, lora_alpha, lora_dropout, target_modules), but nothing consumed them — use_lora was silently ignored and training fell back to the full-parameter LLM path. This PR wires LoRA up for the Qwen3-0.6B LLM inside FunASRNano.

Changes

• funasr/models/fun_asr_nano/model.py: when llm_conf.use_lora is true, replace the LLM target nn.Linear layers (default q_proj/v_proj) with lora.Linear adapters (reusing the existing funasr/models/lora/ module used by Paraformer). The pretrained base weight is shared and frozen; trainable lora_A/lora_B are added. Adapter params are created in the base weight's dtype (bf16), so the LoRA path works without relying on
  autocast to reconcile dtypes.
• examples/.../fun_asr_nano/lora_finetune.sh + docs/finetune.md / finetune_zh.md: LoRA recipe with freeze semantics and deployment (fold) notes.

Design notes

• Pairs with the existing lora_only / mark_only_lora_as_trainable flow already in
  train.py / train_ds.py:
• lora_only=true (default in the example) freezes every non-LoRA parameter → pure-LoRA.
• lora_only=false + unfreeze encoder/adaptor conf keeps them trainable while LoRA-tweaking only the LLM (the recipe used for the medical ASR deliverable).
• Checkpoints save the full state dict (unchanged base + adapter params), so a LoRA run can be resumed / decoded with the same use_lora=true config. No new save/load code.
• Inference needs no merge step (forward adds the adapter output); for a standalone deployment checkpoint, fold with W' = W + (alpha/r) * B @ A and drop the lora keys.

Validation

• Real model build: 56 adapters injected (28 Qwen3 layers × q_proj+v_proj), base frozen,
  LoRA trainable.
• Optimizer step updates only the 112 LoRA params (0 non-LoRA change).
• 2-step smoke training via funasr-train-ds completes; saved checkpoint's lora_A/lora_B are bf16 matching the base dtype; merge folds all 56 pairs cleanly.
• eval()/train() auto merge/unmerge verified (matches the upstream lora.Linear
  convention used by Paraformer).

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.