Support for training lm_head in freeze finetuning_type

[upskyy]

What does this PR do?

Fixes #3694

Before submitting

A ValueError occurs because lm_head is not included in model.named_parameters().
https://github.com/hiyouga/LLaMA-Factory/blob/main/src/llamafactory/model/adapter.py#L124-L131

[rank6]:   File "/llama-factory/src/llamafactory/model/adapter.py", line 135, in _setup_freeze_tuning
[rank6]:     raise ValueError(
[rank6]: ValueError: Module lm_head is not found, please choose from norm, embed_tokens

[upskyy]

If embed_token and lm_head are tied together like gemma2, a KeyError occurs in model.named_parameters() where lm_head cannot be found.

[hiyouga]

We do not need to specify lm_head to trainable if the model ties word embeddings, since the input/output embeddings share the same weights. Just use freeze_extra_modules: embed_tokens