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

[james-willis]

Adds RS_SummaryStats(raster, statType[, band[, excludeNoDataValue]]), Sedona Spark's single-statistic band summary. statType is one of count, sum, mean, stddev, min or max and is case-insensitive. stddev is the population standard deviation. The band defaults to 1 and excludeNoDataValue defaults to true, as in Spark. Over no pixels, count and sum are 0 and the other statistics are NaN, which matches Spark (checked against 1.9.1).

Bit-for-bit with Spark

Spark computes these with Apache Commons Math (DescriptiveStatistics plus a population StandardDeviation). Commons Math's mean adds a second-pass correction term, and its variance subtracts an accum2²/n term. Both change the low bits compared with a naive formula. The kernel repeats that arithmetic step for step, including the left-to-right sum and the NaN-skipping min/max, so results match exactly with no tolerance.

The anchored fixture is six values near 1e8 with fractional parts. On it, every shortcut from Commons Math gives a different answer from the anchors: a different summation order, a mean without the correction pass, a variance without the accum2 term, and a variance around the naive mean. A unit test (fixture_separates_the_shortcuts) checks that each one does, and the anchors are Spark 1.9.1's own output.

Each statistic re-scans the band rather than copying its values out: count, sum, min and max take one pass, mean two and stddev three. The band bytes are already in memory, and a Vec<f64> copy would be 8 bytes per pixel outside the query's memory accounting. count decodes nothing, so it also works on 64-bit bands whose values f64 can't hold.

The semantics follow Sedona's main branch, not the 1.9.1 release. apache/sedona#3366 made a NaN nodata value exclude NaN pixels after 1.9.1 shipped, so that spark-parity case is an xfail until the suite's Sedona pin moves past it. Two other divergences are recorded as xfails: Spark 1.9.1 reads UInt32 pixels as signed, so values at or above 2^31 wrap negative; and SedonaDB rounds a fractional file nodata (0.5) into an integer band's type, which drops real 0 pixels (the known issue test_rs_bandnodatavalue.py already records). A 64-bit integer pixel beyond 2^53 is an error rather than a silently rounded value, as in RS_Value, and the error names the pixel.

Shared pixel scan

The kernel reads the band through pixel_scan.rs, the crate-private whole-band scan and NodataMatcher that landed with RS_MinConvexHull (#1351). Integer nodata compares by bytes and float nodata numerically.

Tests

• Rust unit tests: bit-exact anchors from Spark and the guard that the fixture separates the shortcuts, case-insensitivity, invalid statType, including or excluding nodata, NaN nodata, empty input, a single value, NaN data, band errors, 64-bit precision and counting, and every arity with NULL propagation (including a NULL band).
• python/sedonadb/tests/functions/test_rs_summarystats.py: GeoTIFF fixtures read through RS_FromPath, checked against numpy. The values are chosen so every statistic is exact in binary, so the comparisons use == even though numpy sums pairwise.
• integration/spark-parity/test_rs_summarystats.py: anchored parity against Sedona Spark 1.9.1 over float64, float32, int16 and uint8 bands (34 pass, 3 xfail as described above).

[jiayuasu]

LGTM