Added contact-aided invariant ekf to mujoco_infer example

[Kongx231]

Converted and tested C++ code for Contact Aided Invariant Extended Kalman filter for estimating global rotation, position, and linear velocity.

The filter is implemented in playground/open_duck_mini_v2/mujoco_infer.py. It requires the path to the newly added URDF files for the duck located in playground/open_duck_mini_v2/urdfs/open_duck_mini_v2.urdf.

An example run is:

mujoco_infer.py -o /path-to-repo/Open_Duck_Playground/checkpoints/BEST_WALK_ONNX_2.onnx --urdf_model_path /path-to-repo/Open_Duck_Playground/playground/open_duck_mini_v2/urdfs/open_duck_mini_v2.urdf

Note: I left out the RL implementation changes since I don't have time to clean it up.

TODO

• Filter utilizes Drake for kinematics computations, however, this can be replaced with MuJoCo. What is required is the kinematics from the IMU to the foot for each leg and their spatial Jacobians.
• Implement the filter with JAX (this depends on the previous point since we cannot use Drake in the JAX enclosure). Ideally, we train a policy using the filters estimates as the observations.