[arrow] Fix second-precision timestamp conversion for pre-epoch values

[thswlsqls]

Purpose

fix #9599

• ArrowUtils.nonCastedTimestampToEpoch() computed the epoch second for precision == 0 as timestamp.getMillisecond() / 1000. Java integer division truncates towards zero, so a pre-epoch value with sub-second millis moves forward across the epoch boundary: -500 (1969-12-31T23:59:59.500) becomes 0, i.e. 1970-01-01T00:00:00Z.
• Use Math.floorDiv so truncation goes towards negative infinity. The 1-3, 4-6 and 7-9 precision branches are already negative-safe and are untouched.
• This aligns the write direction with the read direction, fixed the same way in [arrow] Fix nanosecond timestamp conversion for pre-epoch values #9299 (Arrow2PaimonVectorConverter.convertEpochToTimestamp).

Tests

• Added ArrowFormatWriterTest#testWritePreEpochSecondPrecisionTimestamp: an ArrowFormatWriter -> ArrowBatchReader round trip over TIMESTAMP(0) and TIMESTAMP_LTZ(0). It fails on master (expected: "1969-12-31T23:59:59" but was: "1970-01-01T00:00") and passes with the fix.
• mvn -pl paimon-arrow clean install — 71 tests passed, 0 failures.