Separate correlation means from the jackknife patch layout

[cailmdaley]

Full-sample TreeCorr correlation means depend on the jackknife patch layout; seeded patches make a layout reproducible but do not remove that dependence.

• correlation.measure_with_patches runs the published-means pass without patches using bin_slop=min(1, b_target/bin_size). The patched pass omits bin_slop, so TreeCorr uses its standard default for covariance and resampling.
• ξ±, M_ap² and measured ρ/τ share this mechanism, including cross-catalogue and tomographic pairs; no shear_psf_leakage changes. In Derive reporting xi means from an exactly nested fine grid #400, the fine-grid means pass retains bin_slop=1; angle_slop stays at its default.
• Cache sidecars record both configurations, so products from the former shared-tolerance policy do not match. Plot consumers reuse full-precision columns without repeating covariance jobs.
• Synthetic layout tests are bit-exact for published means on one thread while covariance estimates differ. Reverting to patched means fails the ξ± and M_ap² regressions. The SACC producer retains dense jackknife covariance when its text cache exists.
• Full container suite at the stacked Derive reporting xi means from an exactly nested fine grid #400 head, slow tests included: 537 passed; three environment failures reproduce on T (seven missing configured paths and two NFS FITS-cleanup errors). Host workflow suite: 11 passed, 1 skipped.
• SP_v1.4.6.3, 1–250′/20 bins/100 patches: layout-dependent mean difference falls from 0.322σ rms / 0.809σ max to 5.24×10⁻¹²σ rms / 2.58×10⁻¹¹σ max. Independent 12-thread runs are not bit-identical; the residual is consistent with floating-point reduction round-off.
• In the stacked Derive reporting xi means from an exactly nested fine grid #400 reporting check, reusing the fine product, the default-tolerance covariance-only pass took 3.95 min (seeded) / 3.76 min (saved centres) at 8 threads, excluding catalogue loading and patch-centre construction.

Depends on #394 (merge/develop-into-tomo); #400 derives reporting ξ± from a nested fine grid. The effect of changing bin_slop on jackknife covariance remains open in #401.

— Luna on behalf of Cail