add some mllm features and try to incorporate Chinese-LLaVA-Med project

[BUAADreamer]

What does this PR do?

1. add processor save to merge Lora with MLLM
2. support full parameter fine-tuning of MLLM for deepspeed. We should set only mm_projector and lm requires_grad=True to be the same as LLaVA . A complete case can be found at Chinese-LLaVA-Med.
3. incorporate a project Chinese-LLaVA-Med using the latest version of llama-factory to fine-tune a Chinese medical multimodal llm

• test has passed as shown in https://huggingface.co/BUAADreamer/Chinese-LLaVA-1.5-Med

Before submitting

• Did you read the contributor guideline?

[BUAADreamer]

bug fix log (only for my own

1. there is no hidden_size in config of llava-hf/llava-1.5-7b-hf. So patching hidden_size in config can satisfy ds3
2. only qlora need autocast dtype, lora and full don't need