Fix reversed argument order in H1JoystickGaitTracking action-rate penalty

[rootkiller6788]

Bug

In mujoco_playground/_src/locomotion/h1/joystick_gait_tracking.py, the action_rate reward for the registered H1JoystickGaitTracking environment was called with its three arguments in the wrong order:

"action_rate": self._cost_action_rate(
    info["last_act"], info["last_last_act"], action
),

The method signature is _cost_action_rate(self, act, last_act, last_last_act) and computes:

def _cost_action_rate(self, act, last_act, last_last_act):
    c1 = jp.sum(jp.square(act - last_act))
    c2 = jp.sum(jp.square(act - 2 * last_act + last_last_act))
    return c1 + c2

So with the current call, act=info["last_act"] and last_act=info["last_last_act"]: the first-derivative penalty c1 measures the change between the two previous actions and completely ignores the action just taken, and the second-derivative term c2 is computed about the wrong point.

Every other locomotion environment with the same 3-argument helper passes the arguments in the intended order, e.g. g1/joystick.py, go1/joystick.py, t1/joystick.py, op3/joystick.py, berkeley_humanoid/joystick.py:

"action_rate": self._cost_action_rate(
    action, info["last_act"], info["last_last_act"]
),

Impact

H1JoystickGaitTracking has action_rate=-0.01 in its reward config, so this penalty is active during training. Because the arguments are shifted, the penalty signal is wrong at every step: the current action is excluded from the first-derivative term, so the policy receives a mis-scoped action-rate cost.

Fix

Reorder the arguments to (action, info["last_act"], info["last_last_act"]), matching the method signature and every sibling environment.

Verification

• Confirmed the method signature (mujoco_playground/_src/locomotion/h1/joystick_gait_tracking.py:456-462) and that all ~7 sibling envs with the same 3-arg helper call it with action first.
• Confirmed H1JoystickGaitTracking is a registered environment (_envs/_cfgs in mujoco_playground/_src/locomotion/__init__.py) and action_rate=-0.01 is non-zero in its default reward config.
• Confirmed last_act/last_last_act are initialized in reset() and updated each step(), so the bug manifests on every transition.
• Confirmed the bug is still present in current upstream main (fetched via GitHub API).
• py_compile passes on the edited file.