Save best checkpoint to disk instead of keeping it in memory

[KerekesDavid]

This ~halves the memory usage during training :)

Also included a fix for multi-gpu evals crashing when used with wandb.