[datafusion-spark] unix_millis / unix_seconds truncate toward zero instead of flooring for pre-epoch timestamps #26128

Describe the bug

The Spark-compatible unix_millis and unix_seconds functions in datafusion-spark return different results from Spark for timestamps before 1970-01-01 that have a sub-unit component. Spark floors the result; DataFusion truncates toward zero.

For example, one microsecond before the epoch gives unix_millis = -1 in Spark but 0 in DataFusion.

unix_micros is not affected, since it does no unit conversion.

To Reproduce

DataFusion (datafusion-cli --spark):

SELECT
  unix_millis(arrow_cast(-1, 'Timestamp(Microsecond, None)'))        AS millis_a,
  unix_millis(arrow_cast(-1001, 'Timestamp(Microsecond, None)'))     AS millis_b,
  unix_seconds(arrow_cast(-1, 'Timestamp(Microsecond, None)'))       AS secs_a,
  unix_seconds(arrow_cast(-1000001, 'Timestamp(Microsecond, None)')) AS secs_b;

Spark 4.0.0 (spark.sql.session.timeZone=UTC):

SELECT
  unix_millis(timestamp_micros(-1)),
  unix_millis(timestamp_micros(-1001)),
  unix_seconds(timestamp_micros(-1)),
  unix_seconds(timestamp_micros(-1000001));

| input (µs) | function       | Spark 4.0.0 | DataFusion |
|-----------:|----------------|------------:|-----------:|
| -1         | `unix_millis`  | -1          | 0          |
| -1001      | `unix_millis`  | -2          | -1         |
| -1         | `unix_seconds` | -1          | 0          |
| -1000001   | `unix_seconds` | -2          | -1         |

Exact multiples of the unit (e.g. -1000 µs → -1 ms) and non-negative values (e.g. 1500 µs → 1 ms) already match

Expected behavior

Match Spark: floor-divide the microsecond value by 1,000 for unix_millis and by 1,000,000 for unix_seconds.

Additional context

The functions were added in #19892. SparkUnixTimestamp::simplify rewrites the call as CAST(CAST(ts AS Timestamp(<unit>, "UTC")) AS Int64). The Arrow cast to a coarser timestamp unit truncates toward zero, which is wrong for negative values with a remainder.

Spark computes these in TimestampToLongBase (datetimeExpressions.scala) using Math.floorDiv.

I searched existing issues and PRs and didn't find this reported. I have a fix with sqllogictest coverage and will open a PR shortly.