As reported by @ghostdogpr, property based tests involving filtered generators were not terminating in certain circumstances.

The issue was that we need to use slightly different concepts of filtering for generating versus shrinking random values. For shrinking if a value doesn't satisfy the predicate we want to continue to examine its shrinks because even if a value doesn't satisfy the predicate its shrinks might and could be part of the optimal solution. In contrast, for generating random values if the value doesn't satisfy the predicate we just want to discard the entire thing, including its shrink tree. Previously we weren't doing this so if the predicate was _ > 0 and the initial random value was a negative number we would fruitlessly search its shrinks for a value satisfying the predicate.

I also added some documentation on transforming generators generally being preferable to filtering them.