Added support for Apple Silicon GPUs (MPS) alongside CUDA.

Changes:

1. Created utils/device.py with helper functions:

   • get_device(): Returns best available device (cuda > mps > cpu)
   • to_device(): Moves tensors/models to the best device
   • set_default_device(): Sets default tensor type based on device

2. Modified all files that used cuda directly:

   • evaluate.py
   • llama.py
   • main.py
   • train_ppo.py
   • train_rm.py
   • train_sft.py
   • trainers.py

The code now checks for the best available device and uses it appropriately.