feat(array): add Interval extension dtype - #10468

Summary

• Tracking Issue: Epic: Time series data support #10171
• Stacked on feat(array): add Duration extension dtype #10467.

Adds vortex.interval, a builtin extension dtype for calendar intervals of months, days, and nanoseconds, and maps it to Arrow Interval and DataFusion ScalarValue::Interval*. It joins AnyTemporal alongside Duration.

Changes

• Interval ext type in vortex-array/src/extension/datetime/interval.rs: empty metadata, storage Struct { months: i32, days: i32, nanoseconds: i64 } (non-nullable fields), so each part compresses on its own. Values display as ISO 8601 with a sign per part.
• Intervals compare field by field (months, then days, then nanoseconds), which matches Arrow's ordering of IntervalMonthDayNano.
• Arrow: all three interval units import without loss; export is MonthDayNano, or YearMonth/DayTime when requested and no value would lose a part or precision.
• DataFusion scalars convert in both directions. DuckDB and CUDA reject Interval for now.

API Changes

• TemporalMetadata gains an Interval variant.
• TemporalMetadata::time_unit() now returns Option<TimeUnit> (None for Interval).
• TemporalArray::new_interval constructor.

Testing

• Rust: ext type validation/display/scalar round trip; field-by-field comparison; Arrow import of all three units and narrowing export (including lossy-value errors); scalar export; DataFusion scalar conversion.
• Python: pa.month_day_nano_interval() round trip.

Open question

vortex.interval is in no edition yet, the same open question as vortex.duration in #10467.

Follow-up

Map DuckDB INTERVAL (DuckDB compares normalized intervals, so interval comparisons should not be pushed down from DuckDB), push down interval literals from DataFusion, ts ± interval.