Flip interval field ordering (#5654)

[tustvold]

Which issue does this PR close?

Closes #5654

Rationale for this change

There have been reports that the definition we use is backwards, and whilst the specification is kind of ambiguous, we should probably follow what arrow-cpp does.

This is a draft as I think we should look to get integration test coverage in first to make sure that this is actually correct.

What changes are included in this PR?

Are there any user-facing changes?

Aside from the obvious there are a couple of potentially problematic changes that result from this:

• The ordering of intervals changes, as this is an arbitrary order driven by the bit representation
• Data in IPC files will be interpreted differently

The order was already arbitrary so the first may not be too problematic, but the latter is potentially. I do, however, suspect that few people are encoding intervals in IPC files else we would have had reports of this before

[tustvold]

The interval ordering is now even more arbitrary 😅

I suspect this is the most user-visible change, as I suspect not many people are manually bit-twiddling

[tustvold]

Looks like the cast kernels will also need updating, but I'd like to get feedback on if this is even the right approach to take first.

[avantgardnerio]

I think we should look to get integration test coverage in first to make sure that this is actually correct

💯

[pitrou]

You can probably add some PyArrow integration test(s) in https://github.com/apache/arrow-rs/blob/master/arrow-pyarrow-integration-testing/tests/test_sql.py ? Both to _supported_pyarrow_types and perhaps also a type-specific test to check that the values are interpreted correctly.

[david542542]

Thanks for opening this PR. Any idea on ETA for when this would be merged in?

[tustvold]

As this is a breaking change with potentially major downstream implications, I suspect it will likely be a while before we incorporate this PR, if even at all. The major potential pain point concerns IPC data files that have been written or are being sent over the network. We need to be very careful here.

I suspect this will be a timeline in months I am afraid

[tustvold]

Closing in favour of #5769