feat(rust/sedona-raster-functions): add RS_MinConvexHull

[james-willis]

Adds RS_MinConvexHull(raster[, band]), Sedona Spark's footprint of a raster's data pixels. It returns the polygon over the smallest pixel-aligned rectangle of the grid that holds every non-nodata pixel, in world coordinates. In effect it is RS_ConvexHull with the nodata margin trimmed off. Without a band, a cell holds data if any band holds data there. The result is NULL when no pixel holds data, and a band with no nodata value covers the whole grid.

The output matches RS_ConvexHull: a planar geometry with the raster's CRS as an item-level CRS, and the same ring order (upper-left, upper-right, lower-right, lower-left), written through footprint::write_footprint_wkb. On a skewed grid the corners follow the grid axes.

Each band is scanned inward from its edges and the scan stops at the first data pixel from each side: rows down from the top, rows up from the bottom, then columns in from the left and right over the rows in between. A band that is mostly data, the common case, costs a handful of pixels instead of a full scan, and a mostly-nodata band costs at most about two full scans. Without a band argument, bands stop being scanned once the whole grid is covered. The bottom and side scans read the band through reversed and transposed views of the same buffer. A unit test checks the result against a full scan for every data mask of a 4x3 grid.

The semantics follow Sedona's main branch, not the 1.9.1 release. apache/sedona#3366 made an all-nodata raster return NULL and fixed NaN nodata after 1.9.1 shipped. In 1.9.1 an all-nodata raster gives a polygon built from Integer.MAX_VALUE sentinels, and a NaN nodata value matches no pixel, so both spark-parity cases are xfails until the suite's Sedona pin moves past it.

Shared pixel scan

This adds pixel_scan.rs, a crate-private helper that walks every pixel of a 2-D band through its NdBuffer strides, plus NodataMatcher, which compares integer pixels to nodata by bytes and float pixels numerically. The RS_BandIsNoData (#1350) and RS_SummaryStats (#1352) PRs carry the identical file, so whichever merges second rebases without conflicts in it.

Tests

• Rust unit tests: margin trimming, the edge scan against a full scan, a band without nodata, all-nodata and NaN nodata, the band union, band errors (including a negative band), north-up and skewed world placement, and NULL propagation.
• python/sedonadb/tests/functions/test_rs_minconvexhull.py: north-up and skewed GeoTIFFs read through RS_FromPath, checked against rasterio's affine applied to numpy's data-pixel bounds (exact, dyadic coefficients).
• integration/spark-parity/test_rs_minconvexhull.py: anchored parity against Sedona Spark 1.9.1 for each band, all bands, skew, CRS, a band without nodata, a NULL band and an out-of-range band (8 pass, 2 xfail as described above). The skewed case uses exactly representable coefficients: SedonaDB and Spark order the affine additions differently, which can differ in the last bit for ordinary decimal skews (the same holds for RS_ConvexHull).

[jiayuasu]

LGTM