By not including the *T_impl.hh files in the corresponding *T.hh files,
but only providing explicit template declarations and instantiations for
the types we use them for, we can avoid recompilation of code that
(indirectly) depends on the *T.hh files, which turn out to be many.

This still leaves a lot of re-linking with libACG, but at least it's
much faster than before.