Add timestep-aware sparse KD weighting

Force cached KD masks during Stage 3b

Guarantee masked assistant tokens in diffusion training

Implement stage-aware real-run training pipeline

Add comprehensive README with architecture details and training recipe