Custom physics engine beats MuJoCo G1 baseline 10x

This shows humanoid control policies can train dramatically faster than standard MuJoCo setups, which could slash iteration time for robotics research.

I just beat mujoco playground's baseline by 10x now. The policy is up on dingbotics. I wrote a custom physics engine. It still doesn't handle perturbations that well because fable decided to leave some bugs behind