Feat: speedups

[liyier90]

Refactors for speedup

Contains various algorithm refactor/changes to improve execution speed, numbers provided are from cProfile + snakeviz where possible.

dabble.zone_count

• Shift .buffer(1) from contains() to __init__() (~12.8s to 4.47s)
• Create Point from btm_midpoint once (~8.86s to 5.48s)

draw.tag, draw/utils/general.py:project_points_onto_original_image

• Use np.rint instead of np.round since we are rounding to integer anyway. (~5.54s to 4.16s)
• Use points.reshape directly, since all points in PKD are already float. Use np.ndarray * tuple to do column size calculations to avoid inplace modification or having to create a temporary array. (~5.48s to 3.68s)

dabble.tracking, dabble/trackingv1/tracking_files/iou_tracker.py

• Use a mask to filter out detections so the original numpy array can used instead of a list with del. Enables computing iou with vectorization instead row-wise calculations (~23.5s to 18.8s)
• Use np.maximum and np.minimum directly without np.c_ (~18.8s to 7.28s)
• Multiply directly instread of using np.prod to avoid unnecessary reduce operation (~7.28s to 6.02s)

draw.legend

• Perform addWeighted operation on a portion of the input frame, avoid copying the entire frame (~6.57s to 4.28s)

model.posenet (unable to visualize with snakeviz after applying decorator)

• _clip_to_indices use np.rint
• use np.linalg.norm instead of sum squared
• Apply numba.jit decorator

model.efficientdet

• Performs preprocessing calculations in one line (~43.5s to 26.7s)
• Reimplement np.pad with np.zeros to use numba.jit (~26.7s to 10.8s)

Dependencies

• Add numba >= 0.56.4 dependency
• Bump numpy to 1.18.5 for linting requirements
• Remove upper limit for tensorflow (Offending code in efficientdet already removed, tested to be working up to tensorflow==2.11.0)
• Relax torch and torchvision version pinning (tested to be working up to torch==1.13.0 and torchvision==0.14.0)