fix(g1_env): restore gravity compensation feedforward

[MiaoDX]

Summary

This refreshes the gravity-compensation fix from the earlier closed PR #20, but with a clearer reproduction path and explicit regression coverage.

The root issue is that G1Env.queue_action() always sends body_tau = 0, even when enable_gravity_compensation is enabled. In practice this makes the flag ineffective in the active G1 control-loop path.

On our side this showed up most clearly in the grasp pipeline: when running the normal G1 controller + planner flow with gravity compensation enabled, pressing H to return home would let the upper body sag/drop first and only then recover toward the home pose. Restoring gravity feedforward brings that path back to normal.

Changes

• restore gravity-compensation feedforward in decoupled_wbc/control/envs/g1/g1_env.py
  • compute gravity torques from the latest observed joint state
  • map them into actuator order
  • send them through body_tau instead of always sending zeros
• normalize the control-loop default for gravity_compensation_joints
  • when gravity compensation is enabled and no explicit joint list is provided, keep the intended default as ['arms']
  • this avoids passing an explicit None through the config path and silently changing the compensation scope
• add focused regression tests for both behaviors
  • queue_action() produces nonzero torques when gravity compensation is enabled
  • queue_action() keeps zero torques when gravity compensation is disabled
  • control-loop config defaulting preserves ['arms']

Why this is safe

• the torque path is still gated behind enable_gravity_compensation
• when the flag is off, behavior stays unchanged
• the joint-group default matches the existing documented/operator expectation for G1 upper-body compensation
• the change is covered by small unit tests rather than only by end-to-end grasp behavior

Testing

Locally verified with:

PYTHONPATH=submodules/gr00t-wbc uv run python -m pytest submodules/gr00t-wbc/decoupled_wbc/tests/sim/test_g1_env.py -q

Context

PR #20 already identified the missing gravity-torque application, but that branch was later discarded and is now conflicting. This PR supersedes that change with refreshed context, a slightly stronger config fix, and explicit regression tests.

[MiaoDX]

@ZhengyiLuo PTAL.