Add harbor.rl + Tinker RL training example

[benediktstroebl]

Adds a minimal RL training example that uses harbor.rl's step()/grade() interface with Tinker for the training loop.

• train.py bridges harbor.rl environments to tinker's Env/EnvGroupBuilder abstractions
• Downloads Terminal-Bench 2.0 tasks via the harbor registry
• Adapts harbor.rl's BashTool and grade() as tinker-compatible tool and reward functions