Speed up inference by overlapping I/O with GPU work

[kyroX-D]

What this does

Right now the inference loop processes frames one at a time read from disk, run GPU, write to disk, repeat. The GPU just sits there waiting whenever we're doing file I/O, which really adds up on longer sequences especially with 4K EXR files.

This PR fixes that by:

• Reading frames ahead in a background thread (3 frame buffer) so the next frame is ready when the GPU needs it
• Writing output files in a separate thread so the GPU doesn't have to wait for disk writes
• Removing a torch.cuda.empty_cache() call that was running on every single frame, forcing the CUDA memory allocator to constantly free and re-allocate memory for no reason (the actual cleanup already happens when switching between models)
• Tightening up some overly broad exception catches that could accidentally hide real errors

Files changed

• CorridorKeyModule/inference_engine.py
• backend/service.py

Testing

• All 319 unit tests pass
• All e2e workflow tests pass
• The 9 failing tests are a pre-existing OpenCV EXR issue, not related to this change
• Still need to benchmark with a real clip once the model checkpoint is publicly available