Reduce NN inference time

Reduces NN inference time by keeping very similar performance.

Usage of __constant__ memory together with refactor of NNPropagation functions in an inline way, to make it easy for the compiler to use the same compilation unit as much as possible.

This MR also adds a smaller NN (550 params instead of 750) and keeps a similar performance as shown in the plots.

The same treatment is applied to the TwoTrackMVA evaluation.

Some plots to show how the performance is maintained:

(Thanks a lot to @ahennequ for all the advice and help!)

Goes with lhcb-datapkg/ParamFiles!78 (closed)