[Feature request] RL integration of Tinker API into Cua-bench

[dongyuanjushi]

RL integration of Tinker API into Cua-bench

• Customize a Cua using open-sourced model, e.g., OpenCUA-7B
• Integrate RL workflow of Tinker API with the interface to sync model weights and agent relaunch for the customized Cua

[ddupont808]

Feel free to proceed!