Add accuracy tracking to PRs.

[jenskeiner]

CodSpeed benchmarks that run on PRs and compare to the baseline on develop already give us an idea if code changes had any impact on performance. The same is not true for accuracy.

The existing automated tests compute various transformations and compare the results to a reference. If the error exceeds a defined bound, the case fails, and thus ultimately the whole CI run. This way, some inadvertent breaks can be captured.

However, it would be more useful to track changes to accuracy figures more precisely. This could flag degrading or improving accuracy that does not cross a threshold. This is particularly interesting in light of upcoming performance and accuracy work, as in #222.

This PR adds some custom code to add the following:

• Each in-repo PR gets an automated comment that mentions unchanged/improved/regressed tests in terms of accuracy.
• The comment also includes a link to a webpage (served via GH Pages) that includes heatmaps with absolute results (aggregated over cases with varying parameters N and M that should not be driving the error as it is calculated) and relative results vs. develop.
• For PRs only get the comment with the summary. Nothing gets uploaded to GH Pages since Fork PRs run with read-only privileges.

At this stage, this new way of presenting accuracy figures seems to be working ok, although I would imagine there's still a lot of room for improvement. The goal of getting this in now is to gather more evidence that the changes in #222 not only improve performance of the direct 1D NDFT transforms, but also the accuracy by avoiding the accumulation of round-off errors.

[github-actions]

Accuracy report

No develop baseline yet to compare against — showing this PR's absolute accuracy in the full report. A baseline appears once changes land on develop.

📊 Full accuracy report

[codspeed-hq]

Merging this PR will not alter performance

✅ 132 untouched benchmarks

---

_{Comparing feature/accuracy-checks (cfdc22b) with develop (56ea7dc)}