[XLA] Guard AVX, SSE and NEON instructions

On OSX you currently run into linker errors because unsupported
instructions are registered.

clang on OSX doesn't respect __attribute((WEAK)) (see PR#14893).
So we add ifdefs to register only the supported instructions.

Also register __sincos on __APPLE__ platforms.

TEST=Build tensorflow/compiler/aot/tests:tfcompile builds
successfully on OSX (10.13.2)