fix: use MacOS M1 series to train evaluator and inference

[zhudezhong]

Issue

• considering running CRAG on Mac M1 series computers? #12

Improve macOS setup & inference robustness (macOS requirements, MPS guidance, optional vLLM backend)

Summary

This PR improves the macOS developer experience and prevents inference from failing due to missing vllm. It adds a macOS-friendly requirements file, documents Apple Silicon mps usage, makes vllm optional in CRAG_Inference.py (with a Transformers fallback), and adds a repo-wide .gitignore.

Key Changes

• Add .gitignore: Ignore Python caches, virtual envs, build artifacts, uv caches, and common ML outputs.
• Add requirements-macos.txt: Provide a CPU-friendly dependency set for macOS users.
• Update README (macOS section):
  • Recommend using requirements-macos.txt on macOS.
  • Document how to select mps (Metal) when available for PyTorch workloads.
  • Explain why macOS does not rely on vLLM by default (experimental support / may require source build), and that inference falls back to Transformers when vllm is unavailable.
• Make vLLM optional in inference (scripts/CRAG_Inference.py):
  • Remove the hard vllm import.
  • Add --generator_backend {auto|vllm|transformers} (default: auto).
  • Prefer vLLM when available; otherwise fall back to a pure-Transformers generator to avoid ImportError.

Motivation

• requirements.txt includes packages that frequently fail on macOS.
• CRAG_Inference.py previously required vllm, causing macOS users to hit ImportError after following macOS install guidance.
• Apple Silicon users can accelerate Torch workloads via mps.

How to Test

• Install (macOS):
  • pip install -r requirements-macos.txt
• Inference:
  • Default behavior (auto backend) should run without ImportError even if vllm is not installed.
  • Force Transformers backend: add --generator_backend transformers.
  • If vLLM is installed: use --generator_backend vllm.