updated training eval

sorted imports

final runs and update blog

seprated blog from readme and added new readme

refactor: move training code to scripts/, add train/eval split, tune GRPO hyperparams