Add LLaVA LoRA fine-tuning example

[SahilKumar75]

Summary

• add LLaVA-NeXT image feature packing and Mistral text-model compatibility
• add a LLaVA LoRA fine-tuning script that trains language-model adapters while freezing the vision tower and projector
• document LLaVA 1.6 generation and JSONL fine-tuning data format

Testing

• python3 -m pytest llava/test_llava_next.py -q
• python3 -m py_compile llava/llava.py llava/language.py llava/generate.py llava/train_lora.py llava/test_llava_next.py
• python3 -m black --check llava/llava.py llava/language.py llava/generate.py llava/train_lora.py llava/test_llava_next.py
• python3 -m pre_commit run --files llava/README.md llava/generate.py llava/language.py llava/llava.py llava/requirements.txt llava/test_llava_next.py llava/train_lora.py

Closes #605