refactor(cosmo_val): thin the orchestrator — extract survey-stat + pseudo-Cℓ primitives, pin the estimator

[cailmdaley]

What

Thins the cosmo_val orchestrator by moving stateless computation out of the mixin methods into shared primitive modules — the architecture this package is built around (cosmo_val/ = the stateful orchestrator class; top-level modules = stateless primitives that analysis scripts also import directly). Two extractions:

• Survey stats → survey.py. n_eff_density, area_from_coords, ellipticity_dispersion, additive_bias, effective_survey_stats — pulled out of catalog_characterization; the mixin methods are now thin wrappers. (scratch/guerrini hand-reimplements several of these — they can now import them.)
• Pseudo-Cℓ estimator → new top-level pseudo_cl.py. pseudo_cl_geometry, make_namaster_bin, get_n_gal_map, apply_random_rotation, get_pseudo_cls_map, get_pseudo_cls_catalog — the NaMaster estimator, now a reusable primitive module. The mixin orchestrates over it.
• glass_mock unified. It carried drifted copies of the estimator; powspace_bins and get_n_gal_map now route through the shared primitives. compute_two_point_cl[_map] are deliberately kept (they use unweighted fields and return the coupled spectrum, a genuine divergence the primitive's contract doesn't cover) — documented inline.

New test coverage, and what it caught

The pseudo-Cℓ estimator — a Paper II deliverable — had zero numeric pins. This PR adds test_pseudo_cl.py (golden values for all three binning schemes, the occupancy map, and the map/catalog EE·EB·BB spectra). Pinning it surfaced — and this PR fixes — a real reproducibility bug:

• Fixed: the pseudo-Cℓ noise-debiasing was non-reproducible. apply_random_rotation called np.random.seed() with no argument, re-seeding from OS entropy on every call, so the noise realizations differed run-to-run. The fix drops the bare seed; the primitive uses its rng argument and the debiasing + covariance-sampling paths thread a generator seeded from a new CosmologyValidation.cell_seed (default 8192). Same numerical scheme, now reproducible. (This is the one deliberate behavior change in the PR; everything else is structure-only.)
• The catalog path has ~2×10⁻¹² cross-process numerical drift (NaMaster NmtFieldCatalog), so those pins sit at rtol=1e-6; the binning, occupancy map, and map-path estimator are bitwise-identical (rtol=1e-9).

Verification

ruff check src/sp_validation clean; all four modules import (pseudo_cl, glass_mock, survey, cosmo_val). Full non-slow suite: 136 passed, the new pins all pass, value-drift (test_cosmo_val, test_b_modes) green. One non-blocking note: n_eff_density adds a sum_w2 > 0 guard the inline code lacked — identical for any real catalogue, a safety improvement.

Stacked on #207. Draft for review.

— Claude on behalf of Cail