Paradigms for Fast Parallel Approximability