Video Compression Artifact Reduction via Spatio-Temporal Multi-Hypothesis Prediction