dt.round() slow/fast path use different rounding

[mcrumiller]

Checks

• I have checked that this issue has not already been reported.
• I have confirmed this bug exists on the latest version of Polars.

Reproducible example

@MarcoGorelli
This is an example case produced in the hypothesis test for test_dt_round_fast_path_vs_slow_path:

from datetime import datetime
import polars as pl

datetimes = [
    datetime(1980, 1, 1, 0, 0, 0, 0),
    datetime(1969, 1, 1, 0, 0, 0, 1),
    datetime(1969, 1, 1, 0, 0, 0, 2),
    datetime(1969, 1, 1, 0, 0, 0, 3),
    datetime(1969, 1, 1, 0, 0, 0, 4),
    datetime(1969, 1, 1, 0, 0, 0, 5),
]
every = "2us"
s = pl.Series(datetimes)

out = pl.select(
    s.dt.round(every).alias("fast"),
    s.dt.round(pl.Series([every] * len(datetimes))).alias("slow"),
)

print(out)

shape: (6, 2)
┌────────────────────────────┬────────────────────────────┐
│ fast                       ┆ slow                       │
│ ---                        ┆ ---                        │
│ datetime[μs]               ┆ datetime[μs]               │
╞════════════════════════════╪════════════════════════════╡
│ 1980-01-01 00:00:00        ┆ 1980-01-01 00:00:00        │
│ 1969-01-01 00:00:00        ┆ 1969-01-01 00:00:00.000002 │
│ 1969-01-01 00:00:00.000002 ┆ 1969-01-01 00:00:00.000002 │
│ 1969-01-01 00:00:00.000002 ┆ 1969-01-01 00:00:00.000004 │
│ 1969-01-01 00:00:00.000004 ┆ 1969-01-01 00:00:00.000004 │
│ 1969-01-01 00:00:00.000004 ┆ 1969-01-01 00:00:00.000006 │
└────────────────────────────┴────────────────────────────┘

Issue description

The fast path rounds down and the slow path rounds up.

Expected behavior

Both should use same rounding.

Installed versions

main

[mcrumiller]

The different rounding routines are here:

• fast path: https://github.com/pola-rs/polars/blob/main/crates/polars-time/src/round.rs#L39-L41
• slow path: https://github.com/pola-rs/polars/blob/main/crates/polars-time/src/windows/window.rs#L85-L89

Do we want to:

1. round toward zero;
2. round away from zero;
3. round toward the even digit, as most rounding functions do?

(1) is the easiest to implement, (3) is more universally implemented.

[MarcoGorelli]

thanks @mcrumiller for spotting this! yeah this needs fixing

in Series.round Polars rounds away from zero, so perpahs .dt.round should be consistent with that

EDIT: having said, should the user really care about where "zero" (1970-01-01) is? A consistent "round down" approach might be better here?