EH: Finally cloning

[JosephTremoulet]

As an optimization, we'll want to clone (at least some) finally handlers on (at least some) non-exceptional paths. TBD whether this is done in-reader, post-reader in llilc code, or by some general tail-duplication-esque optimization in LLVM. The invocation of the beginfinally intrinsic could be removed from the non-exceptional clone (it represents run-time effects of the 1st pass unwinding).