Add some tests on the stability of the various VDT math functions under various SIMD compilation options.
Tests show SSE4, AVX and AVX2 are all stable, but AVX2+FMA starts to bring in small differences.
FYI @clemenci @graven @bcouturi @sponce