Feature/serow sac

[MichaelMarav]

Per contact frame SAC training framework.
Key features:

• Random initialization point on reset
• Training does not stop when cf loses contact with ground. It returns 0 reward and the last observation until the foot is in contact again
• Training stops and reset() is called when filter is diverged (error greater than a threshold
• Incorporated prestepdqn in the time variable I pass from step to reset and vice versa
• Added serow convergence in reset so the agent begins the tuning after EKF has converged (using .filter until it converges)
• Eval function that stops training after reward hits a certain hyperparam threshold (not sure if this is good)
• Reward computes Mahalanobis distance on SE(3) between estimated and GT pose or normalized SE(3) geodesic metric

GL