Marginalise n(z) uncertainty with modes of an n(z) ensemble (as DES Y6), not shift/stretch #469

Claude (Opus) on behalf of Cail.

Our inference models n(z) uncertainty with a per-bin mean shift Δz: CosmoSIS photoz_bias in cosmo_inference/cosmosis_config/templates/. A shift plus stretch is the usual extension. Both leave the n(z) shape fixed. With a SOM n(z) built from a few spectroscopic fields, the shape is where much of the uncertainty is: tails, spikes from structure in the spectroscopic fields, and correlated changes across bins.

DES Y6 instead takes an ensemble of n(z) realisations, compresses it into a few modes, and samples their amplitudes in the likelihood. The modes capture shape changes and cross-bin correlations that a shift cannot. We should do the same.

Done when:

• An n(z) ensemble exists for a tomographic n(z). For the working SOM n(z) (Working tomographic n(z): SOM calibration in sp_validation, read by cosmo_val as a file #468), this can be bootstrap realisations over calibrators and fields, emitted by the calibration.
• A step reduces the ensemble to its leading modes plus the mean, and writes them where the inference reads the n(z) (SACC or the CosmoSIS nz input).
• CosmoSIS samples the mode amplitudes with their priors in place of photoz_bias. On mocks, the posterior recovers the input when the true n(z) is drawn from the ensemble, and the Δz-only model's bias is quantified against it.

Related: #455 and #468 (where the n(z) is made), #274 (the photo-z group's SOM n(z), whose own ensemble should feed this), #220 and #460 (CosmoSIS with tomography).