Fix compilation failure caused by the special treatment of caml_flambda2_invalid for zero_alloc (added in #2112) in the presence of unboxed types.

The middle end generates caml_flambda2_invalid with return=false and return type set to typ_void. In Selectgen, we sometimes change it to return=true and typ_int, but the context may not be compatible with the return type of int. This should not matter because the code is unreachable, but it triggers an assertion:

>> Fatal error: Illegal move between registers of differing types (I/77[%rax] to F/78[%xmm0])

The proposed fix updates return=false at the end of Zero_alloc_check on all these special primitives.

• For Cfg, we rerun dead code elimination pass if needed. This is expensive but performed only when needed.
• For Mach, we rely on Deadcode pass to eliminate the illegal moves before register allocation. Changing returns field is expensive because we reallocate all functions annotated with zero_alloc. A cheaper alternative (see reverted commit) is to directly change Deadcode pass.

A proper fix would be to let the middle end eliminate dead functions even if they are annotated with zero_alloc assertions (i.e., revert #1378), and instead propagate zero_alloc assertions to instructions, similarly to the way we do for assume using debug info, but currently debug info may not be reliable enough.