fix: support encrypted Qwen vision models and RK picture frames

[athenawxd]

Qwen RKLLM image analysis now loads a CEMC-encrypted vision.rknn through Model Guard using the installed device certificate. Plain RKNN files retain their native loading path, protected-model failures never fall back to native loading, and the adjacent model.rkllm continues to use the vendor loader.

The picture worker also accepts host-backed CPU/Rockchip frames after EnsureHostData succeeds. Previously it additionally required GetHostData, rejecting valid frames whose pixels are exposed through GetData with InvalidParam/13 before inference. Failed host transfers remain rejected.

Validation:

• Reproduced the original failure through authenticated image upload and picture analysis.
• Real picture-worker regression: old code fails; fixed code passes 25 assertions, including failed-transfer and inactive-frame rejection.
• CPU build and full suite: 1,013 test cases, 17,217 assertions passed.
• RK3576 production-release engine with RKLLM and Model Guard built successfully.
• Changed files pass clang-format 18. Full-tree formatting still reports 25 pre-existing files, including dependencies under node_modules.
• Encrypted model initialization was observed on RK hardware. End-to-end inference with the updated picture-worker binary has not yet been verified on the device.