Also clean up the remaining docstrings

Read integer columns through the SDM schema so nullable BIGINTs round-trip
exactly instead of being silently coerced to float64 by pd.read_sql_query.
Fall back from validityEndMjdTai to the legacy validityEnd name for
older APDB fixtures.

The metadata rows and flag-tag overlays are now declarative tables driven
by a shared loop, instead of a long sequence of inline fig.text calls.
A new heights kwarg lets subclasses with extra panels position the rows
themselves.

New compare module: match_catalogs cross-matches two DiaSource frames
within optional group columns, flux_residuals computes per-band/per-flux
residual statistics for matched pairs, and match_to_truth attaches the
nearest truth row to each detection.

display_images is rewritten around a shared overlay-collection helper:
catalogs are loaded once, color-coded by flag group when requested, and
the same overlays are reused for every frame. New display_images_ab puts
two butlers side-by-side, and extract_timestamped_messages flattens an
LSST JSON log.

Three helpers: lightcurve plots a per-band psfFlux history for a single
diaObject (with optional forced-photometry overlay), cutout_grid lays out
science/template/difference cutouts for many sources in a mosaic, and
summarize_run produces a per-visit summary of an APDB run.

Extends PlotImageSubtractionCutoutsTask with a lightcurve panel beneath
the science/template/difference cutouts. The panel shows the diaSource
history for the associated diaObject and overlays forced-photometry
measurements for visits without a detection.

Walks the per-(visit, detector) DiaSource/DiaObject/DiaForcedSource
datasets that DiaPipelineTask writes alongside its APDB inserts and
concatenates them into single DataFrames matching the APDB SDM schema.
Useful for QA without round-tripping through a live APDB.

image_diff_stats reports median, MAD, stdev, kurtosis, and mask-plane
fractions for a single difference image. pixel_compare aligns two
difference images on the same dataId and returns their pixel difference,
ratio, mask XOR, and summary statistics.

subtraction_quality_report consumes a per-detector
*_spatiallySampledMetrics table and prints percentile summaries plus a
three-panel diagnostic figure for diffim_chi2PerPix,
psfMatchingKernel_residualNorm, and dipole_density.

Walks every <task>_metadata dataset in a butler collection, extracts
timing fields via QuantumResourceUsage.from_task_metadata, applies a
per-task threshold (keep the whole task if any quantum exceeds it), and
returns a tidy DataFrame with an optional box plot.

`find_objects_sharing_sources` walks the (diaSource, run-1 diaObject,
run-2 diaObject) graph to return the full connected cluster reachable
from a starting diaObjectId. `classify_association_clusters` enumerates
every such cluster across two runs and labels each as matched, split,
merged, or tangled.

`plot_cutouts_with_object_markers` renders per-diaSource cutouts with
color-coded diaObject markers and "other diaSource" overlays.
`plot_objects_sharing_sources` is a two-column wrapper that compares
the run-1 and run-2 views of the cluster reachable from a starting
diaObjectId.

The existing erase call only erased the image, not any markers drawn on top of it.