Add RL-friendly velocity and effort action commands

[yuecideng]

Description

This PR adds a stable flat Box policy contract that routes action-term slices to typed qpos, qvel, and qf robot commands.

Key changes:

• Support single and mixed control terms on explicit, disjoint joint_ids or control_part selections.
• Validate batched shapes and finite values, then clip physical commands to robot limits by default.
• Hold effort commands across every physics substep during action decimation.
• Preserve the exact raw flat policy action in RL, reward, rollout, trajectory, and dataset paths.
• Add optional tanh-bounded Gaussian policies with the Jacobian-corrected log probability used by PPO and GRPO.
• Keep bare env.step(tensor) as the legacy qpos path while accepting explicit direct command mappings.
• Document action ranges, drive-mode caveats, mixed-control configuration, and bounded-policy behavior.

Compatibility notes:

• A single qpos action term retains its historical tensor return for task reward code.
• Built-in simulator RL enables action squashing automatically only for normalized [-1, 1] manager bounds.
• Custom manager bounds leave automatic squashing disabled and emit guidance.

Dependencies: None.

Issue: None linked.

Type of change

• Enhancement (non-breaking change which improves an existing functionality)

Validation

• black --check --diff --color ./ — 849 files clean.
• Focused Gym, manager, simulation, policy, evaluation, and rollout tests — 128 passed, 1 deselected.
• Standalone shared-rollout regression — 1 passed.
• RL routing, APG, differentiable collector, and trainer regressions — 23 passed.
• Real CPU simulation qvel/qf rollout — 1 passed.
• make -C docs html — succeeded. The build reports 668 existing repository warnings; warning-as-error remains blocked by that baseline, with no new warnings from the changed action pages.

Screenshots

Not applicable.

Checklist

• I have run the black . formatting gate for the code base.
• I have made corresponding changes to the documentation.
• I have added tests that prove the enhancement works.
• Dependencies have been updated, if applicable (none required).