multi_inspiral: avoid repeated power chi^2 calculations in the critical loop

[titodalcanton]

In the presence of loud injections (and perhaps glitches) and with large sky grids, multi_inspiral spends most of its time doing chi^2 calculations. This reduces that.

Standard information about the request

This is an efficiency update.

This change affects PyGRB.

Motivation

This addresses the first thing I described in #5148, namely the fact that some multi_inspiral jobs can take many hours just doing repeated power chi^2 calculations in the critical loop over time slides and sky points. As the calculation is a single-detector operation, it does not depend on time slide or sky position, so its result be cached for reuse in the following iterations.

Contents

I introduce a cache for the power chi^2 values and DoF. Every time a new chi^2 calculation is required, I check which time samples had their chi^2 calculated before, and I only call the chi^2 code for the new samples.

This change, together with #5157, brings the wall clock time from 35 hours to 8 for the test job I have been profiling in #5148.

@spxiwh noted that the point chi^2 code is not necessarily the most efficient when there are many points to calculate; in fact, coh_PTF_inspiral used to do the old FFT-based calculation. Still, I already know that this PR leads to a huge speedup in the limiting cases, so I propose we postpone further optimizations to later PRs.

Links to any issues or associated PRs

#5148

Testing performed

I noticed that power chi^2 values tend to be non-deterministic even across runs of the same code done a few minutes apart, so checking that the results do not change is a bit challenging, but I am working on this.

I also need to verify that this does not slow the code down in the case of well-behaved noise without loud signals in it.

Additional notes

N/A

• The author of this pull request confirms they will adhere to the code of conduct