Switch scopes from keeping decls in SmallSet to keeping them in the new

SmallPtrSet data structure.  This datastructure handles the 'nonsmall' case
quite gracefully, with an efficient exponentially probed hashtable.  This is
important for handling global scope, which gets many thousands of decls (e.g.
every function and enum value).  Of course the typical inner scopes are still
as efficient as ever.

On my mac pro, this speeds up parsing carbon.h from 0.59s to 0.168s (3.5x),
and there is still low hanging fruit :).

For reference, GCC on the same system takes 0.33s for -fsyntax-only.

llvm-svn: 39317