Reusable figure selections for run-to-run comparison (#121)

[bpurinton]

Summary

Closes #121.

When re-processing the same scene with different ASP parameters (the workflow in stereopipeline-quickstart and across our own mapproj-vs-no-mapproj example notebooks), the diagnostic figures silently changed what they showed between runs — making before/after comparison impossible:

• a fresh ICESat-2 request returned a slightly different point set,
• the "best" profile track (_select_best_track) flipped,
• the best/worst agreement segments (_find_best_worst_segments) moved,
• the detailed-hillshade clip boxes were re-selected from the re-processed intersection-error raster.

All of this selection logic is already deterministic given identical inputs — the non-determinism enters only through the re-fetched ICESat data and the re-processed DEM. So the fix is persist + replay, not seeding.

What changed

• asp_plot/selections.py (new): FigureSelections dataclass + YAML read/write + clip-box ↔ pixel-window helpers. Imports nothing from report.py/fpdf, so it's safe from notebooks. bbox_to_pixel_offset rounds to nearest (not floor) so a replayed clip doesn't drift a pixel from floating-point error.
• CLI asp_plot: always writes <report_stem>_figure_selections.yml next to the report; new --reuse_selections PATH flag replays a prior run's ICESat-2 points, profile track, best/worst segments, and hillshade clips. A single run now also resolves the best track once and shares it across the profile + segment figures.
• StereoPlotter.plot_detailed_hillshade: new clip_windows= kwarg (DEM-CRS bboxes); records the drawn boxes on self.detailed_hillshade_clips. Out-of-bounds boxes warn and fall back to automatic selection.
• Altimetry: load_atl06sr_from_parquet() reuses the exact prior points; plot_atl06sr_dem_profile/plot_best_worst_segments accept pinned rgt/cycle/spot + segments=; get_altimetry_selections() reports the choices. _find_best_worst_segments gains a segment_override.

Per the design discussion, this is one umbrella flag (not granular toggles), reuses the parquet points for exact reproduction, stores clips as projected coordinates, and resolves the intra-run best-track inconsistency. All new kwargs default to auto, so existing behavior is unchanged.

Testing

• New tests/test_selections.py (12 tests): YAML round-trip, bbox↔pixel helpers, hillshade clip replay + out-of-bounds fallback, segment-override (+ bad-override fallback), parquet reuse, and get_altimetry_selections bundling. tests/test_imports.py updated.
• Validated end-to-end on the Mars MOC example (planetary, no internet): run A → run B with --reuse_selections reproduces the detailed-hillshade clips exactly (this is what surfaced and fixed the floor-vs-round off-by-one).
• Pre-existing test_altimetry failures in my local env are an unrelated broken PROJ database (DATABASE.LAYOUT.VERSION.MINOR = 5), identical on main; new tests avoid that path.

Version bumped to 1.16.0.

Notes for reviewers

• Docs: added a "Comparing re-processing runs" section to docs/cli/asp_plot.md.
• The schema is human-readable so a user can hand-edit it to force a particular track or clip.