fix: gate foot_clearance_reward by command and body movement to avoid rewarding standing still

[NeoZng]

The original impl of foot_clearance_reward() encourages the agent to stand still, which could maximize the reward (since exp(-0)=1).

This PR add command and current velocity weights to tell whether the robot is staying still or moving.

Wrong implementation of the function is the main reason why the examples in this repo require a huge amount of iterations to just learn to walk. On contrast, examples in Isaaclab official impl takes only about 500 iters with 4096 envs to walk in complex terrain (though it takes base_lin_vel in policy observation)

[NeoZng]

BTW, it is not recommend to use this reward for terrains including pyramid/slope, better to modify the implementation of feet_height_body() .