Add non-CUDA runtime fallback and documentation

[arunshar]

Summary

• add device fallback in training flow: CUDA -> MPS -> CPU
• add non-CUDA attention/runtime fallbacks so validation runs can complete on macOS/CPU
• make dataloader/evaluation device-aware in prepare utilities
• document the new non-CUDA runtime profile and requirements in README

Validation

• uv run prepare.py
• uv run train.py (completed on MPS using reduced runtime profile)

Conversation: https://app.warp.dev/conversation/bd4fddb2-351a-4f25-8ddb-7a4bb3d4dbb9

Co-Authored-By: Oz oz-agent@warp.dev