ntdll: Treat FPU_sig as a void* and memcpy to and from it as needed.
It may be unaligned on certain architectures (e.g. macOS).
We started seeing this intermittently on Intel Macs after !9060 (merged). Since FPU_sig was returning an XMM_SAVE_AREA32*, the compiler assumed it was correctly aligned, and emitted movaps for a copy. If the pointer wasn't aligned, that of course failed. Presumably alignment is dodgy on macOS because we're returning a pointer a few fields into the context struct - __fpu_fcw is preceded by int __fpu_reserved[2].