feat(engine): compact context crops and stable per-event detection crops

[MateoLostanlen]

This PR addresses two issues with the 4K detection crops:

• Reduce RAM usage: the full decoded 4K frame (~23 MB) was kept in the sliding window and in the alert queue (up to 100 entries) only to cut 224x224 crops at send time. We now keep a small JPEG context crop around the detections, with its position in the frame, and cut the final crops from it at send time. Measured on real 4K frames, a stored entry drops from ~23 MB to ~0.14 MB on average, which frees ~0.5 GB in steady state on a 4-pose setup and removes a ~2.3 GB worst case when the alert queue fills during an API outage.
• Stabilize crops across the alert sequence: crops were re-centered on each frame's bbox, so they moved with bbox jitter. Crop boxes are now computed from the union of tracked bboxes, squared and padded, then frozen for the whole event, so every frame of the sequence is cropped at the same location.
• Multi-detection: overlapping bboxes are clustered and each cluster gets its own frozen crop box. A detection that outgrows its box re-anchors to a new one once.

[codecov]

Codecov Report

❌ Patch coverage is 93.15789% with 13 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.83%. Comparing base (438493e) to head (e29027d).

Files with missing lines	Patch %	Lines
pyroengine/engine.py	93.15%	13 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop     #379      +/-   ##
===========================================
+ Coverage    75.32%   78.83%   +3.50%     
===========================================
  Files            6        6              
  Lines          685      841     +156     
===========================================
+ Hits           516      663     +147     
- Misses         169      178       +9     

Flag	Coverage Δ
unittests	78.83% <93.15%> (+3.50%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.