feat: add fastllm worker for high-performance inference

[crawfordxx]

Summary

Add a new model worker backend based on fastllm, addressing #2521.

• New file: fastchat/serve/fastllm_worker.py — follows the BaseModelWorker pattern (consistent with vllm_worker.py, mlx_worker.py, etc.)
• Streaming generation with support for temperature, top_p, top_k, repetition_penalty, max_new_tokens, and stop strings
• Quantization: --dtype flag supports float16, float32, int8, int4
• CPU threading: --threads flag for CPU-bound inference
• Optional dependency: fastllm is imported only at runtime; clear error message if not installed
• Documentation: docs/fastllm_integration.md with installation and usage instructions

Usage

# Launch controller
python3 -m fastchat.serve.controller

# Launch fastllm worker
python3 -m fastchat.serve.fastllm_worker \
    --model-path chatglm2-6b \
    --dtype int8 \
    --threads 8

Test Plan

• Verify worker starts and registers with controller
• Test streaming generation via /worker_generate_stream
• Test non-streaming generation via /worker_generate
• Test with different --dtype options (float16, int8, int4)
• Test stop string handling
• Verify heartbeat registration with controller

Closes #2521