fix some features in llava-style training

[BUAADreamer]

What does this PR do?

1. support pretraining style like LLaVA by only fine-tuning mm_projector (freeze LM)
2. add mllm_pt_demo data with only image-captioning single-turn data, see BUAADreamer/mllm_pt_demo
3. modify target_modules to a regex string to only tune the Lora params in LM and mm_proj in LLaVA-style MLLM.

Usage

1. pre-train: use sft. set tune_mm_proj=true and finetuning_type="full" and dataset="mllm_pt_demo"
2. Lora's usage is the same as previously.

Before submitting

• Did you read the contributor guideline?

[hiyouga]

LGTM