Preserve contour holes before dissolving chunked label polygons - #1299

fixes #1235

_region_props_to_polygon currently converts each contour into a separate polygon, so an interior contour becomes a filled shell rather than a hole. This produces invalid MultiPolygons with nested shells.

When a label spans multiple chunks, the subsequent dissolve can fill that incorrectly represented hole and return a valid but incorrect polygon. If another label occupies the hole, the resulting shapes overlap despite the raster labels being mutually exclusive. These overlaps can cause downstream area calculations to double-count regions.

This change applies make_valid to invalid geometries with multiple contour rings before chunk fragments are dissolved. Single-ring geometries skip the additional validity check. Regression tests cover occupied holes with different chunk sizes and preservation of disconnected components.

Benchmark results

The benchmark uses the complete cell_labels layer from the public 10x Xenium human lung cancer SpatialData object, distributed as xenium_2.0.0_io_spatialdata_0.7.1.zip on the SpatialData dataset page.

The raster contains 17,098 × 51,187 pixels and 162,254 cell labels. It was loaded into memory once and divided into 1024 × 1024 chunks. Each implementation ran once using Dask's single-threaded scheduler. Data loading and geometry validation were excluded from the timings.

Cell IDs were unchanged. The measured runtime difference was +0.64%. These are single-run wall times. Invalid baseline geometries were counted separately and excluded from intersection calculations.

All 22 vectorization tests passed, along with Ruff and mypy checks.

python -m pip install -e . fsspec aiohttp

mkdir -p temp/xenium-vectorization

python - <<'PYTHON'
from pathlib import Path
import zipfile
import fsspec

url = (
    "https://s3.embl.de/spatialdata/spatialdata-sandbox/"
    "xenium_2.0.0_io_spatialdata_0.7.1.zip"
)
destination = Path("temp/xenium-vectorization/xenium-lung-0.7.1")

# Download only the label layers and root metadata using HTTP range requests.
with fsspec.open(
    url, "rb", block_size=8 * 1024**2, cache_type="blockcache"
) as remote:
    with zipfile.ZipFile(remote) as archive:
        for name in archive.namelist():
            if name == "data.zarr/zarr.json" or name.startswith("data.zarr/labels/"):
                archive.extract(name, destination)
PYTHON

set -o pipefail

python -u temp/xenium-vectorization/benchmark_xenium_vectorization.py \
  2>&1 | tee temp/xenium-vectorization/benchmark.log