PolarsCursor truncates timestamp fractions wrongly for years outside 0000-9999 #1201

Summary

PolarsCursor truncates the fraction of a timestamp to the wrong digits when its year is not written with four digits.

_to_datetimes() (pyathena/polars/converter.py) cuts the timestamp text at a fixed length from _TIMESTAMP_TEXT_LENGTHS (pyathena/converter.py), such as 23 characters for ms and 26 for us, which assumes a four-digit year. Athena writes a year after 9999 with a + sign and five or more digits, and a year before 0001 with a - sign, so the cut drops fractional digits that the time unit holds. The values are returned without an error.

_to_datetimes() converts the timestamp columns of a CSV result read in chunks, and of a whole read whose native parse fails.

Reproduction

Athena writes these values in the CSV result file and the GetQueryResults rows (checked with a live query):

SELECT
  date_add('year', 9000, CAST('2020-01-02 03:04:05.123' AS TIMESTAMP(3))),
  date_add('year', -2021, CAST('2020-01-02 03:04:05.123456789012' AS TIMESTAMP(12)))
-- "+11020-01-02 03:04:05.123", "-0001-01-02 03:04:05.123456789012"

_to_datetimes() returns, with Polars 1.39.0 to 2.0.0:

Expected behavior

The fraction is truncated to the time unit regardless of the width of the year: .123 for ms and .123456 for us.

polars.Expr.str.to_datetime() with the format "%Y-%m-%d %H:%M:%S%.f" already parses these years and truncates fractional digits beyond the time unit, on Polars 1.39.0, 1.44.2, and 2.0.0, so the text does not need to be cut.

ArrowCursor shares _TIMESTAMP_TEXT_LENGTHS, but pyarrow does not parse a signed or five-digit year at all, so its read fails with ArrowInvalid before the cut matters.