[Draft] Alternative rqa trend implementation respecting actual time stamps

[meggart]

Just to collect the results of our experiments with @felixcremer. This is mostly for documentation purposes and does not necessarily need to be merged since the resulting algorithm is 20x slower than the current version.

[felixcremer]

As mentioned at today's meeting. It might be worthwhile to test the tile wide performance with this branch and teat it against the faster implementation

[felixcremer]

I am wondering whether we could improve the speed of this implementation by precomputing the x mean and x squared mean, because this should be the same for all pixels.
From what I understood is the main difference between these algorithm, that with the time series aware version we have to actually compute the x mean and with the standard version that is much easier to compute.