Compute ball velocity vector change in each frame

[huffyhenry]

To speed up alignment, it would be useful to discard Tracab frames that we can be sure don't correspond to Opta events. Otherwise, the N-W matrix is going to have on the order of 1500 x 100'000 ~ 150M entries...

If we make the assumption that during an on-the-ball event (in the sense of Opta), the ball changes its velocity vector, then the Tracab frames where that happens can be found by differentiation of the ball trajectory. These candidate frames can then even get padded a little from both sides and still reduce the size of the problem considerably.

[huffyhenry]

We're still going to align all frames, at least by default, but we will take the (partial) derivative(s) of ball velocity into account.