0.2

• cpu or disk activation offloading, as well as a "hybrid" offloading method that lets you keep up to x MB of activations on GPU, then past that up to y MB of activations will move to CPU, and anything past that will be moved to disk. sort of overflowing to whatever is fastest whenever needed.
• unsloth fixed gradient accumulation
• rouge, google_bleu, and meteor evaluation metrics
• overall code cleanup
• rework dataset config method