ngmix: give the PSF observation a finite weight map (#749)

[cailmdaley]

Closes #749

What

make_ngmix_observation built the PSF observation weightless
(psf_obs = Observation(psf, jacobian=psf_jacob)), while the galaxy
observation right below it carries a full weight map. With no weight, ngmix
defaults to unit weight on the unit-flux PSF stamp, so the galaxy
GPriorBA(0.4) prior handed to the PSF fitter swamps the (tiny) likelihood and
the recovered PSF shape collapses toward the prior (zero ellipticity).

This PR gives the PSF observation a finite flat weight map
psf_wt = 1/PSF_NOISE**2 (PSF_NOISE = 1e-5, the esheldon/aguinot value
Axel's #749 reproduction used) — the PSF analogue of the galaxy weight built
just below. Weight only: PSFEx/MCCD models already carry a little noise, so no
explicit stamp-noise injection is added (Axel's guidance on #749). The galaxy
prior is left in place (option B from the thread) — once a real weight is
present the likelihood reasserts and the prior no longer bites.

Closes #749

One root, two issues

The single weightless line was the root of both #749 (the galaxy shape prior
prior-dominates the PSF fit, collapsing the recovered ellipticity) and the now
-retired #774 (the small finite-weight noise budget of the pre-v2.0
esheldon/aguinot code was dropped in the rewrite). They are the same bug from
two angles; the finite 1/psf_noise**2 weight is the substance of the old
noise-injection, so adding the weight resolves both. #774 carried no separate
deliverable and has been retired as redundant.

Evidence — end to end through do_ngmix_metacal / average_original_psf

A sheared-PSF stamp (psf_shear = (0.05, -0.03), HSM moment truth
g = [+0.0477, -0.0286]) pushed through the real production path; the
*_psf_orig column is the original-PSF fit:

build	g_psf_orig	frac recovered
weightless (bug)	[+0.0001, -0.0001]	0.00
weighted (this PR)	[+0.0477, -0.0286]	1.00

g_psf_reconv ≈ 0 in both builds — the reconvolution kernel is round by
construction, independent of this weight (and is the column the pre-#761
g1_psfo_ngmix name was mis-sourcing).

• Calibration is provably untouched. Every metacal-type shear
  (noshear/1p/1m/2p/2m) is bit-for-bit identical (Δ = 0) between the
  unit-weight (pre-fix) and weighted builds, and the fit is flat across
  PSF_NOISE ∈ {1e-4, 1e-5, 1e-6}. Metacal fits this same psf_obs with a
  prior-free AdmomFitter (MetacalFitGaussPSF._do_psf_fit), which is
  insensitive to a flat weight's absolute scale — so the diagnostic weight
  never reaches calibration. It is the galaxy prior on the diagnostic fit
  (average_original_psf), not weight-independence, that makes *_psf_orig
  sensitive while the calibration is not.
• Weight-only vs weight+noise (ngmix: restore (or justify dropping) the PSF noise-injection / PSF-observation weight #774's question). Injecting explicit
  PSF-stamp noise (σ = 1e-6, 1e-5) on top of the weight moves g_psf, T_psf,
  and R by ≲ 1e-6 — explicit injection is justifiably dropped.

The faint-end reduced-χ² / star-response coupling #774 also speculated about is
tracked separately under the χ²-anomaly investigation (#769), not here.

Tests

tests/module/test_ngmix_weight_validation.py gains three guards: a structural
check that psf_obs carries the finite weight (never unit weight), the #749
end-to-end recovery contrast through average_original_psf (weighted recovers
to 1e-3, weightless collapses below 10%), and the bit-for-bit metacal-shear
invariance.

— Claude on behalf of Cail