Support extract on intervals

[nrc]

This is an early draft PR to demonstrate an Arrow change (see below for details).

Which issue does this PR close?

Closes #6327.

What changes are included in this PR?

This PR simply allows Interval types to be passed through to Arrow. It will require an Arrow upgrade to work since it depends on apache/arrow-rs#6071 (currently Arrow is patched for the CLI in this PR, this will only work if you have modified Arrow in a sibling folder to DataFusion, I'll remove this and update the PR when there is an Arrow release).

Are these changes tested?

Tested in Arrow. Tests to come here.

Are there any user-facing changes?

The extract function will now work with intervals.

[nrc]

apache/arrow-rs#6071 has landed. It just missed the 52.2 release of Arrow, so hopefully will be in 53. Once that is released (scheduled for September), I'll resurrect and rebase this PR and add some tests.

[samuelcolvin]

@nrc can we add

                    Exact(vec![Utf8, Duration(Second)]),
                    Exact(vec![Utf8, Duration(Millisecond)]),
                    Exact(vec![Utf8, Duration(Microsecond)]),
                    Exact(vec![Utf8, Duration(Nanosecond)]),
                    Exact(vec![Utf8View, Duration(Second)]),
                    Exact(vec![Utf8View, Duration(Millisecond)]),
                    Exact(vec![Utf8View, Duration(Microsecond)]),
                    Exact(vec![Utf8View, Duration(Nanosecond)]),

here, then convert durations to Intervals — presumably MonthDayNano is easiest.

[nrc]

These are added to this PR, Arrow PR coming soon

[nrc]

Arrow PR: apache/arrow-rs#6246

[nrc]

After some testing, I found a couple of bugs due to the seconds function making assumptions about date_part which don't hold for Interval/Duration. Calling date_part to get the nanos from an Interval or Duration will often return Null due to overflow, which will cause date_part/extract to return null, even if we should have ignored the nanos and just used the whole seconds (e.g, extracting seconds from a 1000000 second duration should return 1000000, not null). Furthermore, extract(second from interval '1 second') would return 2 because the one second was queried as seconds and nanoseconds.

Both those bugs are fixed, but because the formatting of Intervals has changed with the Arrow update, it is difficult to write tests, so those will come after the proper Arrow update.

[nrc]

Closed in favour of #12514