feat: add CUDA availability check with helpful error message

[tykoo-chen]

Problem

Currently train.py assumes CUDA is available and will crash with a cryptic error if no GPU is detected:

• torch.cuda.get_device_capability() throws an exception
• Users on Mac or without NVIDIA GPUs get confusing error messages

Solution

Add an early check for CUDA availability before any CUDA operations, with a helpful error message that:

1. Clearly states an NVIDIA GPU is required
2. Points Mac users to the MLX forks listed in README
3. Suggests checking CUDA drivers on Windows/Linux

Code Change

# Verify CUDA is available before proceeding
if not torch.cuda.is_available():
    print("Error: No CUDA GPU detected. This script requires an NVIDIA GPU.")
    print("If you're on Mac, see the MLX forks listed in README.md")
    print("If you're on Windows/Linux, ensure CUDA drivers are installed.")
    exit(1)

Benefits

• Better user experience for people trying the repo on unsupported hardware
• Faster path to the correct solution (forks for Mac, driver installation for Windows/Linux)