With this change, we are now correctly lowering `sub_with_overflow`
for signed and unsigned integers of register-sized integers (32-
or 64-bit precisely). We also match LLVM's behavior and so, the
condition flags we now set are:

* unsigned:
  - `add_with_overflow`: `hs`/`cs` (carry set)
  - `sub_with_overflow`: `lo`/`cc` (carry clear)
* signed:
  - `add_with_overflow`/`sub_with_overflow`: `vs` (overflow)