When executing statements in parallel there is a race where we preparethe type info queries multiple times, and so insert into the type infocaches multiple times. This resulted in any existing cached `Statement`to be dropped, running its destructor which attempts to take out thestate lock that is already being held, resulting in a deadlock.Fixesrust-postgres#772.

There is no reason for the buffer and typeinfo caches to share the samelock. By splitting them it means we a) get slightly better performance,but more importantly b) it makes it harder to accidentally deadlock.