Fix Packet*l for 32-bit builds.

Apparently _mm_cvtsi128_si64 and _mm_extract_epi64 are only available on x86_64 targets. Added work-arounds that extract values via bit-casts to double.

I have no idea why only those two instructions don't seem to be available, but all the other *epi64 functions are fine. But with this change, it now builds on both linux with gcc/clang and on windows with msvc when compiling for 32-bit.