fix(rust/sedona-raster): align RS_Rotation with Sedona Spark's formula

[james-willis]

RS_Rotation parity coverage that turned into a kernel fix, in three commits:

1. Probe-first coverage of the divergence: SedonaDB computed atan2(-skewX, scaleX) where Sedona Spark computes acos(scaleX / sqrt(scaleX² + skewY²)), negated when skewY > 0. They coincide only while |skewX| == |skewY|.
2. One parity module per RS_ function: test_rs_scalar.py → test_rs_bandnodatavalue.py, RS_Rotation cases into test_rs_rotation.py (the suite's layout convention going forward).
3. The fix: GeoTransformEx::rotation now uses Spark's formula — the angle of the grid's actual column-axis direction (scaleX, skewY).

Why align rather than document the divergence

• The old formula didn't just diverge under shear: it misreported a plain rigid rotation of non-square (2x3) pixels as −0.7137 for a true −π/6 (the new formula recovers −π/6 exactly, for any pixel shape).
• RS_GeoTransform (feat(rust/sedona-raster-functions): add RS_GeoTransform composite accessor #1221) ports Spark's decomposition, so after this change its thetaI agrees with RS_Rotation by construction — no internal inconsistency.
• acos incidentally fixes a zero-sign nit: the old atan2 returned IEEE -0.0 for axis-aligned rasters where Spark returns +0.0.

Blast radius: GeoTransformEx::rotation feeds only the RS_Rotation UDF; rust unit tests in geo_transform.rs, affine_transformation.rs, and rs_geotransform.rs are rewritten around derived cases (axis-aligned asserts the positive zero's sign; sheared expectation matches Spark bit-for-bit in probing).

The parity tests (all passing, anchored)

• Axis-aligned (north-up, south-up) → exactly 0.0, sign included.
• Rigid ±30° rotations, square and non-square pixels → round(-theta, 12).
• Shear → round(-acos(scaleX/hypot(scaleX, skewY)), 12).

Rotated/sheared queries compare through ROUND(..., 12), with the evidence in a comment above the tests: neither runtime promises a bit-exact acos (Java Math.acos is specified to within 1 ulp of correctly rounded; Rust f64::acos is the platform libm with no accuracy contract), and a 72-angle sweep on macOS vs Temurin 17 measured 2 angles differing by exactly 1 ulp — π/6 among them — with the unlucky set being libm-dependent, so a raw comparison would be cross-platform flaky rather than strict. The anchored constants sit thousands of ulps from a 12-digit rounding boundary, so ROUND collapses 1-ulp pairs without ever splitting them. Axis-aligned cases compare raw and exact.

Verification

cargo test -p sedona-raster and -p sedona-raster-functions green locally; integration/spark-parity at 28 passed, 14 xfailed (RS_Rotation contributes 7 anchored passes, no xfails). This branch carries the #1217 lane so CI runs the parity suite.

Aside, noted while probing and out of scope: SedonaDB evaluates -0.0 = 0.0 as false in SQL where Spark says true.

[james-willis]

drivey by: clean up the parity harness to more comprehensively model 1 file for 1 function

[paleolimbot]

Thank you!