QualityMonitoringLarIqaConfig

No-reference ML-based image quality assessment. Four model architectures are supported; pick one via the oneof. All produce a single quality score on a normalised [0..100] scale. The Spectrum preprocessor takes care of the per-model resize + normalisation; multi-input models (LAR-IQA) get one tensor per scale built behind the scenes. DBCNN/MUSIQ/LAR-IQA are ImageNet-normalised still-image models; UVQ (Google Universal Video Quality 1.5) is a video-native model run per-frame at full 1920x1080. sample_every_ms: minimum media wall-clock (ms) between inferences — the analysis cadence, invariant to the source frame rate. 0 = every frame (latency-gated). Absent (OptionalFloat presence) = the per-model server default (still-image models 200ms; UVQ 1000ms, being video-native at full 1080p).