ImzML recalibration

[LachlanStuart]

• [x] Copy across current recalibration & alignment code
• [x] Either try to improve & package MSIWarp for integration, or copy across the necessary parts
• [ ] Save recal stats, sum spectrum, TIC image to DB as Diagnostic blobs
• [ ] Add feature-flagged UI

[LachlanStuart]

This has been shelved as we found it didn't provide a compelling argument for adoption in its current state. (~15% of good-quality and ~100% of bad-quality datasets got worse results, overall improvement usually less than Raphaël La Rocca's implementation). We decided that it should be re-evaluated with METASPACE ML FDR (#797) once that's available, as the inclusion of mass error in the MSM score is likely to provide a clearer advantage.

My notes, if/when this gets picked up again:

• The most important outstanding shortcoming is explainability. The end-to-end diagnostics in evaluation.py need to be re-tooled to plot mass errors (& stddevs) on a ppm scale, and track peaks through the pipeline by peak index instead of m/z so that a reliable before & after view can be shown.
• Unanswered question: Is it possible to detect and abort if recalibration seems to worsen the results?
• There's a bug that causes excessive memory usage somewhere - seems to be late in the pipeline, possibly while writing the ImzML file.
• Several as-yet undocumented ideas for how to further improve the algorithm:
  • Find recalibration targets based on distinct spectra instead of sum spectra. The spectrum summation step is the most likely culprit for why many known peaks didn't work at all for recalibration in the spotting datasets.
  • Likely reason for underperformance compared to Raphaël's algorithm: that algorithm pre-filters outliers, so RANSAC has less work to do.