Scalar NaN and signed-zero sort keys split WindowGroupLimit rank peers

[sunchao]

Describe the bug

Native WindowGroupLimitExec can drop rows that Spark retains when a scalar FLOAT or DOUBLE ordering key contains different NaN representations or both zero signs. This is a follow-up to #4870.

Spark considers all NaNs equal and greater than non-NaN values, and treats -0.0 and +0.0 as equal. Arrow row encoding compares their raw total-order representations instead. Consequently, native RANK / DENSE_RANK cutoffs can split one Spark peer group. Native sorting also needs to agree: with a secondary ordering key, sorting by raw NaN payload or zero sign can make equal peers noncontiguous.

Steps to reproduce

On Spark 3.5 or later, enable native execution and spark.comet.exec.localTableScan.enabled=true, keeping spark.comet.exec.strictFloatingPoint=false.

Build a local table with runtime NaNs, for example java.lang.Double.longBitsToDouble(0x7ff8000000000001L) and java.lang.Double.longBitsToDouble(0xfff8000000000002L), plus finite values and a partition containing both zero signs. Avoid a Parquet round trip that can canonicalize NaN payloads. Filter a query of this shape to its top rank:

SELECT * FROM (
  SELECT *, RANK() OVER (PARTITION BY p ORDER BY v DESC NULLS LAST) AS rnk
  FROM floating_window_peers
) WHERE rnk <= 1

A regression fixture retaining eight rows in Spark returns only four in unpatched Comet. This reproduces for both FLOAT and DOUBLE with native Sort, Window, and WindowGroupLimit operators. Adding a secondary ordering key also exposes the sorting inconsistency.

Expected behavior

Scalar floating-point sort keys, window peer comparisons, and native range-partition keys/boundaries should use consistent Spark-compatible normalization. Only comparison keys should change; returned values must preserve their original NaN payloads and zero signs.

Additional context

The existing strict-floating-point fallback remains an opt-out. Floating-point values nested in arrays or structs are a separate compatibility limitation.