Clang 3.8 does not seem to have the intrinsic `_mm512_castps128_ps512()`, used
in `vectorf128.h`:1063 for `approx_recipr()` if avx512f is available, but
neither avx512er nor avx512vl are available.  All currently existing
processors have either avx512er (phi) or avx512vl (server-skylake).  So do
specify the latter to work around this.