GPU power QR detection/decoding

[yarikoptic]

currently it is way too slow and typhon has a reasonable GPU we can use.

https://github.com/ReproNim/qr_decode_bm?tab=readme-ov-file#benchmark-results provides initial benchmarking results with claude code. Need to be analyzed etc...

[vmdocua]

Notes from recent discussion we had:

• We use opencv-python to decode video frames from mp4. This used by both nosignal and qr-parse tools. It's not fast process and measured in video length terms/timings +-. So it definitely can be optimized by using GPU @ typhon.
• The qr-parse utility also uses pyzbar.decode API to find QR codes in frame, it's also slow. OpenCV provides own QRCodeDetector API. Investigate if it can be faster than pyzbar and use GPU capabilities.
• Optionally. The nosignal utility uses custom algorithm to check only 6x8 (48) pixels against reference rainbow frame. It's quite fast, and I believe no optimization will be necessary, but in case this will be bottleneck, consider also GPU functionality to match frame.

[yarikoptic]

FWIW, the current claude performed comparison https://github.com/ReproNim/qr_decode_bm?tab=readme-ov-file#benchmark-results best result is over twice the real time for QR detection:

Configuration Time FPS QR Codes
OpenCV native + QRDet GPU (2 workers) 163.4s 145.8 46