Add PP-OCRv5 (multilingual OCR: detection + 7 v5 recognizers)

[Viet-Anh NGUYEN (vietanhdev)]

Adds pp_ocrv5 — PaddleOCR v5 text detection + recognition for 7 v5 languages (Chinese/Japanese, Korean, English, Latin, Cyrillic, Thai, Greek), Apache-2.0, as a collection model. The source is ONNX, so the components override serialize() to provide the ONNX graph directly to AI Hub (same approach as pi05) instead of torch-tracing.

AI Hub validated on Samsung Galaxy S25 Ultra (compile + profile): float TFLite, 100% NPU — det 1.54 ms, ko_rec 4.39 ms, zh_rec 10.84 ms. Accuracy on ReCTS real signage (n=200): 48.5% full-string recall, 66.0% per-char recall. int8 (w8a8) quantization with a calibration set further reduces latency to ~1.2/1.0/0.5 ms.

perf.yaml/release-assets.yaml omitted (scorecard/release CI generates them).

[Kory Watson (kory)]

Hi Viet-Anh NGUYEN (@vietanhdev), thank you for the contribution! While we appreciate your work, at the moment we are unable to accept new models from the community directly. I unfortunately won't be able to merge it for the time being, but I'll keep the PR open.

The team is actively working on several API changes needed for us to start taking community contributions--please stay tuned for that. We'd be excited to keep you in the loop when we make those changes. In the meantime, feel free to upload compiled assets to our Hugging Face community here:
https://huggingface.co/qualcomm-ai-hub-community