More Supervision, Less Computation: Statistical-Computational Tradeoffs in Weakly Supervised Learning