Kalman Filter backward pass could use a direct navigator

[andiwand]

As discussed in the last parallelization meeting, the backward pass of the Kalman Filter should use the existing trajectory from the forward pass to navigate backwards. This is computationally more efficient as we skip the navigation procedure. At the same time we disallow finding a different surface sequence in the backward pass.

This is also a good opportunity to decouple the forward and backward pass and potentially the extrapolation step.