Optimize search loop

[larskuenkel]

This PR optimizes the search loop, by

• Using shared memory for all arrays used in the search loop
• Use njit for performing the harmonic summing. In my earlier test of njit I only used it to gather the values needed for the sum but when I also use it for performing the sum, I see a performance boost.

I am not how much memory pressure this now creates, compared to the previous scheme, so I am not sure how the overall gain is but it should be a bit faster, especially when using more cores.

[larskuenkel]

Some benchmark plots where I run the search step. I tested for a spectrum with max DM ~1600 which would be processed in our curretn scheme with 30 threads.

I test the new scheme against the old scheme and just using shared memory.

sps-compute1 tests




sps-compute4 tests

The new method takes about ~30-50% of the time of the old method.
Interestingly enough, especially for the newer node, using more threads will increase used memory bandwidth while only improving the processing time marginally. This may mean, that we want to cap the the maximum number of threads in some upcoming PR.

I should not that there were not many actual detections in this spectrum. So the time measured here is mostly just performing the harmonic summing, while further processing is limited.