open-pi-zero

use torch.compile by default in training and eval

rename flow schedule `gamma` to `beta`, add file structure to README, remove optimizer cpu offloading