Return efficient strategies for simple filter predicates, instead of rejection sampling

[spockz]

Can we do something like HypothesisWorks/hypothesis#2701 for creating more efficient generators while retaining the ability to write filters so we don't have to worry about rejecting too many inputs?