SIMD ML inference backend

Fast ML inference with SIMD for simple models for things like ghostprob, probnns. Goal is to replace TMVA like inference with hardcoded weights and improve speed at the same time. Loading models only from ParamFiles.

Related to &10 and faster alternative to e.g. ONNXRuntime (see e.g. Rec!3380 (closed)), which should be for more general (selection) applications. Related to Rec#460

Inspired on one hand by https://gitlab.cern.ch/lhcb/LHCb/-/blob/master/Kernel/LHCbMath/include/Kernel/TMV_utils.h, which is more specific to TMVA. But this one would be more close to PyTorch like model building.

current main application in Rec!3729 (merged) and Moore!3026 (merged)

see also application in Rec!3664 (merged); and pipeline (with QMT) in Moore!2768 (merged)

taken pytorch cmake config from !4187

libraries getting can be switched off, see lhcb-core/lcg-toolchains!150 (merged)