Evaluation refactor

Improvements to the evaluation code, as well as new tests to ensure that mAP scores match the pycocotools implementation.

Change list:

• Moved all metric computation code to a deeplabcut/core/metrics folder (as metrics are computed with numpy)
• Cleaned metric computation code so the prediction/ground truth matching always happens
  • Refactored in a way such that no OOM errors should occur, even on very large datasets (>60k images)
• Multi-animal RMSE: only compute RMSE using (ground-truth, detection) matches with non-zero RMSE
• Add compute_detection_rmse to compute "detection" RMSE, matching the DeepLabCut 2.X implementation
• Fixed the bug for PAF models documented in Evaluation error with PAF heads: ValueError: matrix contains invalid numeric entries #2631