Move GSF KL distance to using vec.h Minor additions to vec.h

This is the reason for / follow up to !35341 (merged)

So let me mention @ssnyder and @amorley

KLGaussianMixtureReduction changes

• The major change is to move the code in KLGaussianMixtureReduction.cxx from using directly x86-64 intrinsics to compiler (gcc/clang) vector types (e.g look https://clang.llvm.org/docs/LanguageExtensions.html#vectors-and-extended-vectors) as provided by the wrappers in vec.h e.g the vec type and the relevant methods.

• It seems that the code becomes indeed a bit more readable if we avoid the _mm . But this is also for the reviewers to judge.

• My benchmarks so far indicate same performance (as should be)

vec changes

• Add vmax, as we added vmin in the previous iterations seems good to have both

• Modify the vbroadcast . The reason for this is the clang seems to do a consistent job in making the loop go away (https://godbolt.org/z/zeE6b9, https://godbolt.org/z/G7KsjE) but gcc less so, in O2 ( https://godbolt.org/z/rGsz5W, https://godbolt.org/z/jvK5GY). The new attempt v= x - VEC{0}; produces the same assembly for both compilers (when we have/want the vec types) in the cases I looked at.

• Then there is an annoying fact that the clang version in mac/xcode is not the same as llvm. So 11 in xcode is 9 llvm (https://en.wikipedia.org/wiki/Xcode#Xcode_7.0_-_11.x_(since_Free_On-Device_Development). This in principle is not a huge issue but can be annoying if people try to build AnalysisBase for mac ... so lets try to protect for it.