Add tile-cut stitching follow-up to calculate_tiling_qc

[timtreis]

Summary

Builds on #1157. Adds a follow-up pass that recovers cells which segmentation tiling broke into 2-4 pieces by detecting facing cut edges across tile boundaries and assigning each candidate pair a transparent geometric score. Worst case (4-tile corner) is handled.

Two public functions:

• squidpy.experimental.tl.stitch_tile_cuts(sdata, labels_key, ...) -- reads is_outlier=True cells from the QC table, extracts cut-edge candidates via bbox-edge alignment, scores each pair with a transparent geometric composite (see below), and assembles confident pairs into 2-4-piece groups via union-find with corner-junction validation. Writes 4 .obs columns to the existing QC table -- stitch_group_id, is_stitched, n_pieces, stitch_confidence -- plus a .uns["tiling_stitch"] audit trail (params, score formula, run summary). The labels element is never mutated.
• squidpy.experimental.im.make_stitched_labels(sdata, labels_key, ..., merge_strategy="sum", inplace=True) -- opt-in materialisation of a stitched labels element via a lazy dask LUT, plus a collapsed AnnData with one row per unique stitch_group_id (unstitched cells pass through unchanged, stitched groups collapse). Numeric .obs columns and .X aggregate via merge_strategy (sum/min/max/mean/median/first or callable); group-invariant + non-numeric columns always take "first". Preserves .uns, .var, and any user-added obs columns.

calculate_tiling_qc re-runs now warn and drop stale stitch columns when the QC table is overwritten.

How stitch_confidence is computed

For each candidate pair the four geometric / shape-quality features below are averaged into a single score in [0, 1]. No coefficients are fitted or shipped -- the formula is the entire model and is recorded in .uns["tiling_stitch"]["score_formula"].

feature	what it captures	range
iou	1-D intersection-over-union of the two cut-edge extents along the boundary	[0, 1]
endpoint_match	how closely the chord endpoints coincide -- true cuts share endpoints, unrelated cells don't	[0, 1]
merge_compactness	4*pi*A / P^2 of the union mask after morphologically closing the seam gap. Real cells are reasonably compact; false merges produce weird perimeters	[0, 1]
merge_solidity	union mask area / convex hull area. Real cells are convex-ish; false merges have concave joins	[0, 1]

stitch_confidence = (iou + endpoint_match + merge_compactness + merge_solidity) / 4

A gap_score is also computed but only used as a hard filter (already inside max_gap by construction); it does not enter the score. The two merge_* features are computed by materialising a tight crop around the union of the candidate pieces, closing the gap with a disk(3) structuring element, and running regionprops on the largest connected component.

min_confidence is a threshold on this mean; 0.7 is the default starting point. Tune for your data -- the score is heuristic, dataset-independent, and not a calibrated probability. Review false positives / negatives via make_stitched_labels and the visual test fixture, then adjust.