Use core::arch vector types on FFI

Currently, vector types like x86's __m128d are bind as follows:

pub type __m128d = [f64; 2]
// - incorrect alignment: missing #[repr(align(16))]
// - incorrect code generation: not #[repr(simd)]

It would be better for these types to just map to their definitions in {core,std}::arch::{$ARCH}::__m128d; since that would solve both issues (note that arm, aarch64, ppc64le, .. each has its own different types).

When C functions use these types in their return type, the correct function signature is produced (e.g. fn foo() -> __m128d), but when C functions use these types in argument position, the types are replaced with a pointer to its element type (e.g. fn foo(x: *mut f64); instead of fn foo(x: __m128d).

Using these types on C FFI requires unstable and the repr(simd_ffi) feature. Not using simd_ffi for this is probably undefined behavior. Therefore, trying to wrap any such type on stable should hard error anyways.

cc @emilio

No files or tests are named. Start by tracing bindgen's vector-type handling and generated FFI argument bindings, then compare the output with the relevant core/std arch definitions across architectures. Done means vector arguments use the correct arch types and stable builds reject unsupported SIMD FFI use.