Better Velo clustering, better SciFi cuts

This MR homogenizes some of the logic in the SciFi, tweaks the cuts and introduces a big refactor of the Velo clustering.

• Velo clustering now follows a much cleaner updated logic based on uint64_t.
• SciFi windows are not shrinked down, they are kept as calculated initially.
• collect_candidates now starts from the candidates in the middle and iterates left and right in a "pendulum" fashion.
• add_missing_x now reuses the precalculated window.
• Cuts have been adjusted all around to remove fakes more efficiently in the SciFi. This leads to about -8% fakes, at the cost of about 1% in efficiency.
• This MR leads to a performance boost of about 5% (back to 60 kHz in RTX 2080 Ti).