Migrate cosmo_val + Snakemake to the SACC writers

[cailmdaley]

Make SACC the sole native output format of the cosmology-validation pipeline. Every two-point product (ξ± coarse and fine, pseudo-Cℓ, COSEBIs, pure-E/B, ρ/τ, n(z)) is written through the sacc_io builders into single-statistic SACC part files, and the Snakemake workflow assembles those parts into one terminal SACC file per catalogue version, {version}.sacc. No bespoke writer remains on the data-vector path.

Desired end state

• Each cosmo_val statistic's write step is replaced by a call to the matching sacc_io builder, producing a part file that holds exactly one statistic plus the self-describing n(z) tracers it references. A producer writes a part with a single save(s, path) call; it never assembles a terminal file and never emits a bespoke format alongside.
• A dedicated Snakemake assembly rule reads the parts, concatenates them in canonical order, attaches the BlockDiagonal covariance (analysis blocks plus the dense per-pair fine block), and writes the single terminal {version}.sacc. The terminal file is a pure gather; anything that wants only one statistic targets the corresponding part rule, not a second terminal file.
• Part filenames are DAG internals: consumers bind parts through the workflow's path helpers, never by hardcoding names. {version}.sacc is the only stable name.
• Blinding rides the DAG here. Blinding is per part, at birth (PRD: SACC data format + Smokescreen blinding #241 §4; CLIs and the assembly-time assertion are delivered by Smokescreen blinding wiring (fork protocol, three theory backends, custody) #252): a blind-init rule fires once per catalogue version, each blindable part rule (coarse ξ±, fine ξ±, pseudo-Cℓ) is followed by blind-part so only blinded parts persist on disk, COSEBIs/pure-E/B are derived from the blinded fine ξ± (born blinded), and the terminal assembly asserts blind_commitment is identical across parts. This issue authors the rules; Smokescreen blinding wiring (fork protocol, three theory backends, custody) #252 supplies the machinery.
• The bespoke .txt/custom-FITS/.npz writers on the data-vector path are deleted; nothing downstream reads them.

sacc_io is a fixed dependency and the single source of truth for every builder signature. Metadata keys required on every file: catalogue_version, sp_validation_version, created, npatch.

Acceptance

The workflow produces {version}.sacc per version from parts, with only blinded blindable parts persisted and the commitment assertion exercised; the deleted bespoke writers have no remaining readers; the assembled file matches the sacc_io layout contract.

— Fable on behalf of Cail.

[cailmdaley]

Notes on terminal-file contents and size, from the #245 review discussion — relevant when this PR decides what the final gather contains.

Size budget (5 bins, 15 pairs, production grids)

The data vectors are trivial; the covariance sets the cost.

Component	Points	Covariance on disk
Analysis vector (ξ± 20θ, pseudo-Cℓ EE/BB/EB 32ℓ, COSEBIs n≤20, pure-E/B, ρ/τ)	~5,300	~220 MB dense; ~tens of MB block-diagonal
Integration-grid ξ± (1000 θ × 2 × 15 pairs)	30,000	~0.5 GB block-diagonal; ~10 GB if dense

Two consequences:

1. assemble_covariance now stores covariances block-diagonally (one FITS table per block, Σ block² on disk instead of dense N²; sacc's concatenate_data_sets preserves the blocks through merge). This removes most of the storage cost of keeping products in one file. In particular, Cℓ BB/EB and ρ/τ are cheap (~17 MB even with full cross-polarisation covariance) — no storage reason to exclude them.
2. The integration-grid ξ± is the only expensive component. It is an intermediate: consumed by the COSEBIs/pure-E/B derivation before the gather exists. Recommendation: drop it from the terminal {version}.sacc and let it persist as its per-part file — Snakemake provenance covers traceability. That keeps the terminal file at tens-of-MB scale with everything scientifically load-bearing inside.

Other pieces relevant here

• Covariance variants (OneCov / NaMaster / …): a SACC file holds one covariance, so variants are separate merged files reusing the same per-part data vectors — cheap under block-diagonal storage. Declaration/naming of variants belongs to this PR.
• Flexible containers: the sacc_io writers are being relaxed so optional components (e.g. Cℓ EB) need not be present; consumers select what they need and fail loud on absence. The gather logic here should assume partial files are legal.
• Integration grid size is a parameter, default 1000 θ bins (1k-vs-10k comparison in the B-modes paper found no substantial difference; we may push higher for final production).

— Claude (Fable) on behalf of Cail.