Enhance reward shaping and add tail/arm actuators for T-Rex and Raptor

[kuds]

This pull request introduces significant enhancements to the T-Rex and Velociraptor simulation environments, focusing on biomechanical realism and reward shaping for reinforcement learning. The most important changes include new actuators for tail and forelimb control, expanded and normalized reward components, and updates to tests to reflect these improvements.

T-Rex Environment Improvements

• Added three new tail actuators (tail_1_pitch_act, tail_1_yaw_act, tail_2_pitch_act) for active balance control, increasing the action space from 18 to 21. [1] [2]
• Introduced multiple new reward components: posture, nosedive, gait symmetry, smoothness, heading alignment, and lateral velocity penalty. All rewards are normalized for physical plausibility, and total reward calculation now includes these components. [1] [2] [3] [4] [5] [6]
• Added termination condition for excessive forward pitch ("nosedive") beyond natural lean.
• Updated tests to check for expanded action space, new reward keys, and new termination reasons. [1] [2] [3]

Velociraptor Environment Improvements

• Changed forelimb joints from ball to two hinge joints for more realistic arm movement, and added four forelimb actuators. Also added a third tail actuator, increasing action space from 17 to 22. [1] [2]
• Updated keyframe and actuator initialization to match new joint structure and actuator count.
• Revised tests to validate new action space and actuator count, and updated reward component checks. [1] [2] [3]

Simulation Consistency

• Updated keyframe and actuator initialization in XML files to reflect new actuators and ensure correct simulation resets. [1] [2]

These changes collectively improve biomechanical fidelity and reward shaping, enabling more robust and realistic reinforcement learning experiments.