[NN CF] Momentum vector usage by neural network cluster finder

[ChSonnabend]

No description provided.

[ChSonnabend]

@davidrohr
Here we can communicate a bit on the implementation of the momentum vector usage created by the cluster finder neural network. This works so far, but I need to make the things nicer and faster still. My main points of discussion:

• Is the implementation correct in theory / what needs to be adjusted?
• The current implementation is just work in progress and a bit ugly. Will fix that over the course of the work on it
• Can we somehow avoid increasing the size of native clusters in case the momentum vector is not used? Is there something like ClassDef that we could use here? Is there a better way to decrease the size? So far I already scaled it to int16_t to reduce disk / memory usage, but maybe there's a smarter way...