Get rid of EIGEN_HAS_AVX512_MATH workaround.

The workarounds using EIGEN_HAS_AVX512_MATH seem to have primarily been present to avoid packet ops depending implicitly on prsqrt, which now has a generic implementation in terms of preciprocal and psqrt.

The workaround in PacketMathFP16.h is redundant because such old compilers never had support for AVX512FP16, which is required for that file to be included.

I verified that the intrinsics needed for psqrt/sqrt are indeed available for the older compilers excluded by the workaround.