fix: reintroduce Swin model with MAX ensemble strategy for robust AI art detection

upgrade: switch to state of the art Deep-Fake-Detector-v2 model

fix: refactor deepfake ensemble, fix audio inversion bug, update video aggregation

fix: generate linear spectrogram for audio_spectrogram instead of reusing mel-spectrogram

fix: move model preloading to background thread to prevent event loop blocking

feat: Major accuracy upgrades (EfficientNet, DCT, Temporal, Dual Audio)

A+ Production Grade Performance, Concurrency, and Auth

perf: add background preloading of models at startup to fix cold start delay

Update api.py

Update api.py

Upload 18 files