feat: add GGUF/GGML model worker using llama-cpp-python

[crawfordxx]

Summary

Add a new model worker (fastchat/serve/gguf_worker.py) that loads quantized GGUF/GGML models via llama-cpp-python, enabling FastChat to serve locally quantized models with reduced memory requirements.

Closes #2410

Changes

• New GGUFWorker class extending BaseModelWorker, following the same patterns as mlx_worker.py and vllm_worker.py
• llama-cpp-python is an optional dependency -- only imported when using the GGUF worker
• Supports streaming and non-streaming text generation
• Supports standard generation parameters: temperature, top_p, top_k, max_new_tokens, repeat_penalty, presence_penalty, frequency_penalty
• GPU layer offloading via --n-gpu-layers (set to -1 to offload all layers)
• Configurable context length via --context-len
• Controller registration and heartbeat

Usage

# Install dependency
pip install llama-cpp-python

# Start the worker
python3 -m fastchat.serve.gguf_worker \
    --model-path /path/to/model.gguf \
    --model-names my-model \
    --controller-address http://localhost:21001 \
    --n-gpu-layers -1

Test plan

• Install llama-cpp-python and download a small GGUF model
• Start controller, then start gguf_worker pointing at the model
• Verify worker registers with controller (/list_models)
• Test streaming generation via /worker_generate_stream
• Test non-streaming generation via /worker_generate
• Verify heartbeat keeps worker registered
• Test with --n-gpu-layers -1 for GPU offloading