Refactor training policy registry

[pearlq12345]

Summary

• add a dedicated roboclaw.embodied.policy registry for trainable policies
• move policy config definitions out of CommandBuilder
• replace hard-coded TRAIN_POLICY_TYPES with registry-driven discovery
• let each policy provide policy-specific training args via extra_train_args()
• add tests for policy registration, lookup, and unknown policy errors

Testing

• pytest -q tests/test_policy_registry.py tests/verification/test_preflight.py
• Web UI smoke test on http://127.0.0.1:8765/training
• confirmed the policy list renders correctly
• confirmed training can be started from the web UI with a local v3.0 dataset
• confirmed the training job is created and enters running