ngmix: config-selectable UberSeg neighbour masking (BLEND_HANDLING)

[cailmdaley]

Closes #776.

What

Adds a config-selectable UberSeg neighbour treatment to the ngmix module, alongside the current noise-fill. Additive and default-unchanged: a new option BLEND_HANDLING = {noisefill, uberseg} (default noisefill) selects how a neighbour sharing a galaxy's postage stamp is handled.

• noisefill (default) — the historical behaviour, byte-for-byte unchanged: flagged pixels are replaced by a noise realisation and kept at inverse-variance weight.
• uberseg — hard-masks (weight → 0) every stamp pixel closer to a neighbour's segmentation footprint than to the central object's, leaving the image untouched.

Why

The 2026-06-23 tomography call flagged the current noise-fill as suspect: including blended (FLAGS=2) objects produces B-modes, while discarding them is a strong, non-metacalibratable selection bias. UberSeg (Sheldon's MEDS/ngmix treatment) is the principled alternative the group has long wanted. The deliverable is the option plus the A/B evidence, so the production default can be chosen on data, not priors — this PR is the option; the validation and the default flip are downstream (see #769-adjacent discussion and the linked issue).

How

• uberseg_weight(weight, seg, object_number) reimplements Erin Sheldon's meds._uberseg.uberseg_tree nearest-segment Voronoi partition with a scipy cKDTree. Reimplemented, not depended on: the astronomy MEDS is esheldon/meds (the PyPI meds is an unrelated package), and import meds drags the full fitsio/esutil stack for a single C tree. The surviving central core is a single connected, roughly circular region — emergent geometry of the partition, faithful to MEDS get_uberseg (no separate aperture).
• Plumbed ngmix_runner → Ngmix → do_ngmix_metacal → make_ngmix_observation → prepare_ngmix_weights; Postage_stamp gains a per-epoch segs list. Unknown BLEND_HANDLING fails fast at construction.
• Tests (tests/module/test_ngmix_uberseg.py): synthetic two-object stamp asserts the mask geometry (neighbour-side zeroed, central core single-connected, matches brute-force nearest-segment); the noisefill default is byte-identical to the legacy noise-fill on a shared RNG; uberseg leaves the image untouched while zeroing neighbour/flagged weight. Existing test_ngmix / test_ngmix_weight_validation pass unchanged.

Not in scope (deliberately)

• The segmentation-map source is gated — see the linked issue. UberSeg needs the seg map with object IDs at each stamp; ShapePipe carries only a per-epoch binary flag today, and the seg map with NUMBERs is the coadd's while ngmix fits per-epoch stamps (MEDS interpolates the coadd seg per-epoch). The two options (extend vignetmaker to cut + reproject a coadd SEGMENTATION vignet, vs approximate from the binary flag) are posed there for the experts. This PR implements everything not gated; selecting uberseg without a seg map raises with a pointer to the issue. This PR is therefore not production-complete until the seg source is decided and wired.
• Image treatment under metacal — per the design intent (and Axel's "leave neighbour objects on the image"), uberseg zeroes the weight without noise-filling the image. metacal shears in Fourier space using all pixels regardless of weight, so residual neighbour flux can leak through the shear; the noise-fill-vs-uberseg A/B B-mode comparison is what measures this, and Axel's weight-symmetrisation follow-on (out of scope) is what would address it. Fabian validates the science before any default flip.
• No change to the noise-fill path, the production default, or the get_noise-vs-sigma_mad noise-estimate question (ngmix: the reduced-χ² anomaly on stars (error model) #769).

Part of the ngmix v2.0 epic (#762).

— Claude, on behalf of Cail