Fix sqrt(2) precision, add pre-coregistration, and correct variogram plotting in uncertainty propagation

[dshean]

Fixes and tests for elevation-uncertainty propagation (the uncertainty module), building on GlacioHack#759.
These came out of testing _propag_uncertainty_coreg / _infer_uncertainty on the Longyearbyen sample
DEMs and on 0.5 m data. Defaults are unchanged throughout, so existing results are bit-identical unless
the new options are used.

Changes

1. precision_of_other="same" was a no-op. The √2 "same precision" correction (Hugonnet et al.,
   2022, Eq. 7-8) was commented out in _infer_uncertainty and never threaded down, so for two datasets
   of similar precision (e.g. DEM–DEM) the inferred per-pixel error was ~√2 too large, inflating the
   propagated translation uncertainty. Re-enabled in the heteroscedasticity step (matching the legacy
   dem.py path). Default "finer" is unchanged; an unsupported value now raises instead of silently
   behaving like "finer".

2. Error structure was inferred from un-coregistered inputs. _propag_uncertainty_coreg ran
   _infer_uncertainty once on the raw inputs, conflating true error with the misalignment/terrain
   (and, in the test case, glacier-change) signal. Added an opt-in precoreg flag: when True,
   co-register once, then infer the error structure and simulate on the aligned residual (this automates
   the manual fit→apply→propagate workaround). With precoreg=True, summary["mean"] becomes the
   residual transform (≈0), which is documented. Default precoreg=False preserves current behaviour.
   (Kept opt-in because apply's edge regridding can make NuthKaab diverge on very small extents.)

3. Variogram plotting. _infer_uncertainty returns the correlation function, which renders
   incorrectly when passed to plot_variogram (which expects a variogram). No code change is needed —
   the variogram is reconstructible from the already-public get_variogram_model_func(params). Added a
   test of the γ(h) = sill·(1 − ρ(h)) invariant that underpins the correct plotting recipe.

4. Monte Carlo robustness. A single non-converging simulation aborted the whole run. Diverged
   simulations are now skipped with a warning, with a guard that requires a minimum number of successful
   simulations before reporting. The success counts are also exposed as summary.attrs (nsim,
   n_success, frac_success) so callers can gate on the skip fraction — a high fraction is itself a
   conditioning signal (ill-posed geometry, e.g. flat terrain), so it should be machine-readable rather
   than only visible as a logging warning (402c389).

Tests

tests/test_uncertainty/test_uncertainty_notebook.py (cropped sample data, small nsim): √2 scaling,
invalid precision_of_other raises, precoreg equivalence + determinism, diverged-sim skip, and the
variogram-plotting invariant.

Notes

No public API change: _propag_uncertainty_coreg is private and precision_of_other was already
documented; precoreg / error_applied_to docstrings added.

Disclosure: prepared with AI assistance (Claude). I have reviewed and understand all changes and take
responsibility for them.

[dshean]

Hi @rhugonnet. I see others are testing this branch. Are you planning to review and merge these on your branch before it goes upstream?