[choreo] Cache trajectories read from disk

[Nsl106]

[choreo] Cache trajectories read from disk

[oh-yes-0-fps]

I tested with adding this, it adds complexity and memory overhead semantics with little practical gain. I can reimplement it in java to benchmark if you want.

The thing is you most likely only read each traj one time per match and during development you can wait the extra few ms.