feat: support Spark 4.1 TimeType (time-of-day)

[LuciferYang]

Summary

Spark 4.1 introduces TimeType, representing time-of-day (00:00:00.000000 to 23:59:59.999999) stored internally as nanoseconds since midnight. Lance already supports Arrow Time32/Time64 types at the core level; this PR adds read, write, and filter pushdown support in the lance-spark connector layer.

Since TimeType only exists in Spark 4.1+, and the base module must compile against older Spark versions, all TimeType references use reflection. This follows the same cross-version compatibility pattern already established in the project — e.g., Float2Vector (Arrow 18+) uses class name checks rather than direct imports to avoid compile-time dependencies.

Changes

New files

• TimeUnitAccessor.java — read-path accessor for Arrow Time vectors
• TimeTypeRoundtripTest.java — E2E roundtrip test in the Spark 4.1 module

Write path

• LanceArrowUtils.toArrowType maps TimeType to ArrowType.Time(NANOSECOND, 64) via class name check to avoid compile-time dependency
• LanceArrowWriter adds TimeNanoWriter — Spark stores nanos internally, Arrow Time(NANOSECOND) expects nanos, so the value is passed through directly with no conversion

Read path

• LanceArrowUtils.fromArrowField adds an ArrowType.Time branch that resolves the Spark type via TimeUtils.resolveSparkTimeType() — returns TimeType on Spark 4.1+, falls back to LongType on older versions
• New TimeUnitAccessor handles all four Arrow Time vectors (TimeSec / TimeMilli / TimeMicro / TimeNano) and normalizes them to nanoseconds for Spark. Time32 vectors return int and Time64 return long with no common base class, so a TypedGetter functional interface is used for typed access to avoid per-row boxing overhead
• LanceArrowColumnVector dispatches the four Time vector types to TimeUnitAccessor

Filter pushdown

• FilterPushDown.isFilterSupported now checks filter value types and rejects LocalTime values. The Lance planner layer does not yet support the TIME type, so time predicates are evaluated as post-scan filters on the Spark side
• compileValue includes a defensive LocalTime branch — currently unreachable, kept for forward compatibility when Lance upstream adds TIME support

Tests

• FilterPushDownTest: LocalTime filter rejection covering comparison operators, In, Not/Or/And composites, and processFilters bucketing
• LanceArrowUtilsSuite: ArrowType.Time ↔ Spark type bidirectional mapping
• LanceArrowColumnVectorSuite: read-path assertions for all four Time vector types, verifying normalization to nanos
• TimeTypeRoundtripTest (4.1 module): end-to-end write-read roundtrip