Use arrow cast for hand-rolled temporal unit conversion loops #5090

What is the problem the feature request solves?

Several hand-rolled per-row loops implement temporal unit conversions that arrow's cast kernel performs bit-identically (verified against the arrow-cast 58.4.0 source):

• native/core/src/parquet/cast_column.rs (cast_timestamp_micros_to_millis_array / _scalar): unary(|v| v / 1000) plus manual timezone re-attachment. Arrow's Timestamp-to-Timestamp cast performs the same truncating division and attaches the target tz.
• native/spark-expr/src/conversion_funcs/temporal.rs (NTZ arm of cast_date_to_timestamp): per-row (d as i64) * 86_400 * 1_000_000 into a builder. Arrow's (Date32, Timestamp(Microsecond, _)) cast computes exactly (x as i64) * MICROSECONDS_IN_DAY.
• native/spark-expr/src/utils.rs (millis-to-micros): arity::unary(|v| v * 1000). Arrow's timestamp-unit cast performs the same multiply.
• native/spark-expr/src/datetime_funcs/date_from_unix_date.rs: manual Int32-to-Date32 reinterpretation plus a scalar match arm. Arrow dispatches (Int32, Date32) to a zero-copy reinterpret cast; the scalar path can use ScalarValue::cast_to.

Describe the potential solution

Replace each loop with arrow::compute::cast_with_options to the target type.

Additional context

Caveats to preserve:

• micros-to-millis: safe when source and target tz are both Some or both None. If the source is NTZ and the target has a tz, arrow wall-clock-adjusts values, so that combination must keep the current relabel behavior.
• Date32-to-Timestamp: only the target_tz.is_none() branch is replaceable. The session-tz branch implements Spark's DST ambiguity/gap rules (earlier occurrence, pre-transition offset), while arrow's adjustment returns null on both, so it must stay.

Found during an audit of native code that replicates existing arrow-rs kernels.