Sequential detection fixes and optimizations

[Elricfae]

🎯 Smart Temporal Tracking System — Replacing Fast-Path Batch Processing

Overview:
This PR introduces a temporal memory-based face tracking system that intelligently bridges gaps in identity recognition caused by profile angles, occlusions, and motion blur. It replaces the previous "fast-path" and "force_recognition_in_batch" approach with a more robust frame-to-frame continuity mechanism.

Key Changes:

1. Temporal Memory Integration (sequential_detector.py)

• Added _temporal_memory state to track verified faces across frames
• Implements a two-pass filtering strategy:
  • Pass 1 (Strict): Standard ArcFace identity verification via UI thresholds
  • Pass 2 (Rescue): Temporal matching using IoU + embedding cosine similarity (>0.75 threshold) for faces rejected in Pass 1
• Temporal memory resets per frame to prevent identity drift

2. Jump-Cut Protection

• Frames verified faces with cosine similarity checks to prevent sudden identity swaps between frames
• Only accepts temporal matches with >0.75 similarity (indicating continuous physical movement vs. jump cuts)

3. Frame Worker Architecture Fix (frame_worker.py)

• Added is_sequentially_tracked flag to distinguish between faces vetted by sequential detector vs. inline detection
• Temporally-rescued faces bypass strict UI similarity thresholds (set to 0.0) in the FrameWorker
• Ensures rescued faces aren't rejected a second time downstream

4. Removed Batch-Mode Hacks

• ❌ Removed force_recognition_in_batch global flag
• ❌ Removed fast-path single-face bypass logic
• ❌ Removed relaxed-threshold fallback passes
• ❌ Removed batch-specific UI state injection in job_manager_actions.py and video_control_actions.py

5. Early Exit Optimization

• Added fast return path when no faces survive filtering (skips landmark detection entirely)

6. Improved Output Safety

• Refined array casting logic to handle dynamic landmark dimensions (5, 68, 203, 478+ points)
• Better NaN/Inf sanitization with fewer redundant copies

Why This Works Better:

• ✅ Handles occlusions and profile angles naturally via temporal continuity
• ✅ No more UI state pollution from batch processing
• ✅ Cleaner, more maintainable code (removed 200+ lines of batch-specific logic)
• ✅ Temporally-aware identity rescue without compromising performance
• ✅ Works seamlessly for both batch and real-time processing