Fix mixingtypes for g++-11.

The _mm512_broadcast_f64x2 instruction is no faster than the _mm512_broadcast_f32x4 instruction, and is warned about in intel's documentation. There's also a funny interaction within our GEBP kernels with g++-11 that causes the real component to be duplicated to the imaginary component when optimizations are turned on (-O1 or higher). Removing this instruction fixes the issue.

This fixes both mixingtypes_6 and mixingtypes_7 tests for AVX512.