closure api: always extend integers to full word size

[fdopen]

This will fix #441 and related problems on nearly all ARM and MIPS platforms (x86 doesn't seem to be affected because of different calling conventions).

The problems are return results of closures. If integers smaller than word-size are returned, they must be extended to full word-size. But they are currently always zero-extended, never sign-extended (https://github.com/ocamllabs/ocaml-ctypes/blob/0d0c62f5b2432a49c3d0c965b8fa9d0dd50ea857/src/ctypes-foreign-base/ffi_call_stubs.c#L489 ).

The first commit contains test code to demonstrate the issue: callback_returns_int8_t (fun () -> -128) will sometimes return +128.

If one byte is set to -128 and the higher bits are set to zero, the whole word will be read as 128. It's not a problem for the (exclusive) foreign interface, because it's hard-coded that the higher bits are ignored. But if stub code is generated, 128 is the more likely result (it depends on the compiler and its optimization settings).

[yallop]

Excellent -- thanks for tracking this down! Let's hope this is the last promotion-related issue in the libffi code.