"Many fixes"

[Elricfae]

📋 Summary of Changes

This PR introduces critical bug fixes and performance optimizations across multiple processing pipelines:

🐛 Bug Fixes

• Keypoint scaling: Fixed hardcoded centroid (255, 255) with dynamic barycenter calculation to prevent unwanted drift on off-center faces
• Face detection sorting: Replaced linear confidence/area heuristic with non-linear squared-confidence weighting to properly filter bad rotation artifacts
• Manual rotation reversal: Fixed expand=True handling with proper center-crop and corrected inverse rotation matrix math for all coordinate systems
• Face landmark alignment: Added automatic tilt angle detection and correction in fallback mode using eye keypoint positioning
• Mouth alignment: Replaced fragile pixel-top anchoring with statistical boundary anchoring (mean + 1.5σ) for temporal stability
• Tensor race conditions: Replaced .from_numpy().to() with direct torch.tensor(..., device=device) to prevent CPU-GPU synchronization issues
• PyTorch tensor handling: Added safe CUDA→CPU conversion in keypoint degeneracy checks

⚡ Performance Improvements

• CUDA stream removal: Disabled per-worker local streams (fixes VRAM explosion with OnnxRuntime)
• Landmark reuse logic: Intelligent reuse of 203-point landmarks to avoid redundant neural network forward passes
• Pure GPU control flow: Replaced CPU-side conditions with GPU tensor logic to eliminate sync points
• Thumbnail caching: Re-enabled batch thumbnail processing during workspace loading

🔧 Minor Improvements & UI Fixes

• Denoiser seed validation: Corrected minimum valid seed value from 1 to 0
• Border mask default: Enabled by default to prevent black artifacts
• Workspace state isolation: Snapshot control parameters to prevent race conditions during async media loading
• Code cleanup: Removed obsolete tracking logic comment and fixed code formatting