We are now optimized for the use case when there is 1 representative point per detection (i.e. the center of detection box) and uses a fixed distance function that is the Euclidean distance between the tracker's estimate and that point. Making this up to 10 times faster than the original Norfair implementation.
In doing so, we also dropped the use of `past_detections_length`, `distance_function`, and the concept of "detection scores" since we are not using them anyway. Thus make the dependencies only include `numpy` and `numba`.