feat: Add OC-SORT tracker

[omkar-334]

Summary

Adds the OC-SORT tracking algorithm.
Depends on #6950 (rebase this once that PR is merged)

Behavior

• Output contains ACTIVE tracks only; lost tracks coast internally until they recover or age out.
• Observation-centric state per track is grouped in one record; Kalman state stays in a separate dict for batched prediction.
• Defaults follow the reference (Cao et al., CVPR 2023): det_threshold=0.6, iou_threshold=0.3, delta_t=3, inertia=0.2, use_byte=False.

Mechanisms

• OCM (observation-centric momentum): a velocity-direction consistency term added to the first association cost, gated on pure IoU.
• OCR (observation-centric recovery): a second pass matching leftover detections against each track's last observation instead of its Kalman prediction.
• ORU (observation-centric re-update): on re-acquisition after a gap, the filter is rewound to the last observation and replayed along a virtual trajectory to the new one.
• Optional BYTE stage (use_byte): associates low-score detections to still-unmatched tracks before the recovery pass.

Benchmark

Identical synthetic detection streams with known ground truth are fed to every tracker and scored with motmetrics. 300 frames, 3 seeds, averaged. Occluded objects are removed from ground truth while hidden, so identity metrics grade re-acquisition. Higher MOTA and IDF1 are better; lower IDSW is better.

Identity quality (IDF1)

Tracker	linear	occlusion	nonlinear	crowded
getitrack-ocsort-byte	0.977	0.924	0.795	0.832
getitrack-ocsort	0.923	0.876	0.758	0.810
ultralytics-ocsort	0.918	0.883	0.784	0.809
roboflow-ocsort	0.846	0.788	0.667	0.690
getitrack-bytetrack (reference)	0.969	0.794	0.615	0.657

Detection accuracy (MOTA)

Tracker	linear	occlusion	nonlinear	crowded
getitrack-ocsort-byte	0.954	0.942	0.939	0.891
getitrack-ocsort	0.856	0.848	0.848	0.801
ultralytics-ocsort	0.856	0.848	0.849	0.801
roboflow-ocsort	0.733	0.707	0.711	0.636

How to test

Checklist

• The PR title and description are clear and descriptive
• I have manually tested the changes
• All changes are covered by automated tests
• All related issues are linked to this PR (if applicable)
• Documentation has been updated (if applicable)