Add residual design-effect correction and aggregation calibration parameters

[ammarcsj]

Corrects the between-peptide correlation that residual-decorrelation pruning cannot remove, by measuring the excess sibling correlation over the cross-parent shuffle null and applying it as a Stouffer design effect (deff=1+(n-1)*rho).

It is additionally gated on that level's pre-pruning distance exceeding the tolerance, so when peptides are no more correlated than the shuffle null the gate stays closed, rho remains 0.0 and aggregation is bit-identical to before. A single level-wide excess rho is applied per level, clipped at zero once on the level mean rather than per parent, since per-parent means are noisy and clipping each would rectify that noise into a positive bias. Below a total-sample threshold the source switches to raw pre-pruning correlations, because at low replicate counts the survivor rho reads ~0 while the correlation still leaks into the Stouffer sum.

Also adds the calibration parameters this work required, each as its own commit:

• median_normalization — take the median of the between-condition fold-change distribution as the shift instead of choosing between its median and its mode.
• max_peptides_per_protein — replaces the hard-coded 31-peptide cap with a configurable one.
• residual_decorrelation_cutoff_grid — the default grid now extends to -1.0; the negative part is reached only when no cutoff meets the tolerance, in which case the tightest prunes down to min_keep.
• residual_decorr_corr_mode / _cap — bounds how many samples the sibling-correlation estimate may use, so pruning aggressiveness does not grow with sample count.
• median_on_collapse — when pruning collapses a parent to a single surviving child, aggregate via the median of all eligible children rather than reporting the lone survivor. Since median_z does not scale by sqrt(n), no design effect applies on this path, and single-peptide parents are unaffected at any rho.
• per-level diagnostic plots (before/after/null CDFs and the cutoff sweep trace) written when runtime_plots is set.

Behavior changes at default settings

• max_peptides_per_protein defaults to no cap; the previous behavior was effectively a cap of 31.
• The default cutoff grid extends further than before, so pruning can drop more siblings.
• median_on_collapse and residual_deff_correction default to on.

Testing

tests/unit_tests: 185 passed, 1 skipped. Each commit parses on its own, so the history is bisectable.

[mschwoer]

can you please simplify this by splitting into several statements?

[mschwoer]

nevermind, now I got it..
ping me once it is a good time to format the codebase ;-)