Gradient for LowResObservation slower than expected

[pmelchior]

The runtime for gradient updates in LowResObservation is ~ an order of magnitude slower than the forward path. This is unusually slow, so there's probably a bottleneck. This issue is to run a profiler on the a few steps of Blend.fit with a single LowResObservation (i.e. different WCS between model and observation) to find out where the CPU cycles go.

If the culprit is obvious, we should find and replace it with a more performant implementation.

[pmelchior]

One possible clue: when running the multi-resolution tutorial, my machine is heavily swapping. The code only rarely reaches >100% CPU although it is naturally multi-threaded. So, it looks like something is pretty memory inefficient on my Mac (16 GB RAM).

So, possible culprits are

• a single constant piece that is to large to be efficiently operated on
• temporary pieces that are only required for a single iteration and then get discarded again