The tests didn't pass, and fixing them reliably seems impossible.

The previous shrinker always shrunk the inner strategy first.  This
seems to be the wrong thing to do in almost all non-trivial situations
(e.g., generate a vector length and then a pair of vectors of this
length).  The new shrinker shrinks the outer strategy first (e.g.,
tries to minimize the vector length first and then minimizes the entries
in the vectors).

Given the current implementation of shrinking via simplify/complicate,
I had a hard time doing this without cloning the inner strategy before
simplifying it. So there's a breaking change here because prop_flat_map
now requires the closure to produce a strategy that implements Clone.
This could have been avoided by a change in how simplify/complicate
operate, but this has the potential of breaking much more code than
adding the requirement that the inner strategy is clonable.