Integrate L2CS-Net gaze estimation

Merge all other things coming from feature/integration

evaluation: channel ablation script + feature importance LOPO

chore: MLP pipeline, evaluation updates, feature importance, confusion matrices

Fix open-cv GPU support problem (opencv library changed)

Put all the models together (expect UI)

Add ClearML for mlp

reorg folders, add MLP demo

feat: add optional eye model (YOLO/MobileNet) alongside geometry

feat: stage 2 head pose, eye behaviour, MAR/yawn, tighter focus; add torch deps

chore: add template layout and stage 1 face mesh demo