Add GR00T N1.7 training and rollout support

[EAOZONE]

Summary

This PR adds training and rollout support for NVIDIA GR00T N1.7.

• Add GR00T policy configuration and dependency preflight support.
• Expose GR00T-specific training options in the frontend.
• Include the rebuilt frontend bundle.
• Use RTC chunked inference for GR00T and other relative-action policies.
• Configure the RTC execution horizon from the policy’s action-step count.
• Set the inference queue threshold to zero for stable chunk execution.
• Provide a friendly error message when relative-action policies are run without RTC.
• Add test coverage for relative-action rollout configuration.

Validation

• pytest -q tests/test_train.py tests/test_rollout.py — 34 passed
• ESLint on all changed frontend source files — passed
• npm run build — passed
• git diff --check — passed
• Successfully trained a GR00T policy and ran it on a physical SO-101 robot.

Notes

The repository-wide frontend lint reports pre-existing errors in unrelated files. All frontend files changed by this PR pass ESLint.

[nicolas-rabault]

Nice feature, verified all the GR00T/RTC flags against lerobot v0.6.0, they're correct.

Drop the committed frontend/dist/; CI (build_frontend.yml) rebuilds it on main; it just bloats the diff (~1.4MB) and will conflict.

[nicolas-rabault]

Thanks for the thorough follow-up,.

[nicolas-rabault]

Heads up, I rebased your branch onto main and force-pushed. Your six commits are intact, the only changes were resolving conflicts in the committed frontend/dist bundle (all resolved to main's version, so the branch no longer touches dist at all) and one commit wrapping two lines over the ruff limit.