fix(writer): declare the real zone stride so vortex-jni prunes correctly (#418)

[dfa1]

Closes #418. Stacked on #420. The filtered vortex-jni reads in these tests abort the JVM without #420's alignment fix. Merge #420 first; this PR then retargets to main.

Bug

Our legacy vortex.stats zone map holds one zone per chunk (one chunk per writeChunk call), but declared WriteOptions#chunkSize() (default 65 536) as its zone length. Rust reads the declared length as a uniform stride: row r belongs to zone r / zone_len (ZonedReader::zone_range). So a filtered vortex-jni read pruned using the wrong zone's stats:

chunks	filter	vortex-jni before	after
[5,5,5,3], default options	id >= 7	[]	7..17
[4,4,2,4]	id >= 10	[]	10..13

With small batches every row maps to zone 0, whose max is chunk 0's, so everything gets pruned.

Fix

ZoneMapStatCodec.uniformZoneLength(rowCounts):

• every chunk but the last has the same length, and the last is non-empty and no longer: declare that length, so Rust's stride matches our zones exactly and Rust prunes correctly;
• otherwise declare 0, Rust's "no stride" (LegacyStatsLayout::zoned_reader). Rust then reads the column unpruned, which is correct and only slower.

vortex-java's reader doesn't use the declared length for this layout; it places legacy zones on the physical chunks (#419). So nothing changes on our side.

Not done

• Switching to vortex.zoned (what Writer emits the legacy vortex.stats zone-map layout; Rust writes vortex.zoned with aggregate-spec metadata #418 originally asked for): not needed for correctness. vortex.stats stays the more compatible choice for older readers.
• WriteOptions.chunkSize is now unused. Its only effect was this wrong zone length; writeChunk never splits batches. It should either be removed or start splitting batches; that's a separate decision.
• Irregular batches lose vortex-jni pruning (still correct). Writing per-stride zones independent of chunks would recover it, at real cost; only worth it if there's demand.

Tests

• ZoneMapStatCodecTest.UniformZoneLength: 9 shape cases (short tail, short middle chunk, longer tail, empty tail, single chunk, no rows, u32 overflow).
• javaWriter_jniReader_zoneMapped_filterKeepsEveryMatch: three chunk shapes through a filtered vortex-jni read. The two multi-chunk shapes return [] with the previous writer (verified by reverting it). All pass now.
• Full ./mvnw verify is green.

🤖 Generated with Claude Code