Sample replay items from priority-weighted probabilities and compute normalized importance-sampling weights.
