Pint support for top-level functions

[keewis]

This PR tries to get the pint integration tests (see #3594) to pass. To make sure this does not become a giant PR, it is limited to the top-level functions and the work on the tests for DataArray and Dataset will be done in other PRs.

I added pint master to the upstream-dev CI since that way CI finally helps finding bugs. Should I have edited py36-min-nep18 instead?

• Passes black . && mypy . && flake8
• Fully documented, including whats-new.rst for all changes and api.rst for new API

Also, I finally ran into a situation where I have to use assert_allclose instead of assert_equal_with_units. However, assert_allclose does not check for identical units, I think? Would it make sense to copy the code of assert_allclose and add the strict unit checking or would it be better to just use assert_allclose and something like assert extract_units(expected) == extract_units(actual)?

Failing tests list from #3594:

• align: needs more investigation partially blocked by where, IndexVariable and a bug in assert_equal_with_units and / or the tests, also a behaviour decision on align works (but units in IndexVariable still fail)
• combine_by_coords: needs more investigation partially blocked by IndexVariable and a test bug (due to assign_coords with mixed DataArray / array args removes coords #3483) works (but units in IndexVariable still fail)
• combine_nested: needs more investigation blocked by reindex works (but units in IndexVariable still fail)
• concat: needs more investigation works (but units in IndexVariable still fail)
• merge: needs more investigation works (but units in IndexVariable still fail)
• full_like: pint currently does not implement copyto blocked by Expected behavior of np.*_like functions hgrecco/pint#882
• where: needs more investigation unit is stripped because pint does not implement astype and test bug in test_where_dataset works
• dot: pint currently does not implement einsum works

[keewis]

it seems the changes to Variable._getitem_with_mask will make other tests fail. The reason for the changes is that pint will convert every argument of np.where to the first argument's units, which would convert the array to the fill value's units and would be quite unexpected, I think?. Does anyone know what to do here?

[keewis]

the _getitem_with_mask failures are due to bool not defining __invert__, which means it falls back to int.__invert__. I special-cased it, but we could also modify indexing.create_mask to always return a duck array.

[keewis]

the failing matrix multiply tests should be fixed by hgrecco/pint#926.

I think this is ready for review and merge?

[dcherian]

thanks @keewis

Should I have edited py36-min-nep18 instead?Should I have edited py36-min-nep18 instead?

Let's do this once there's a pint release.

[max-sixty]

Thanks for the review @dcherian !

@keewis FYI I am starting a new job so will be lying low on xarray for a couple of months, forgive the temporary pause in speedy reviews. It was great to work together for the past few months!

The whole team is looking forward to working more closely with you in the future. We'd like to invite you to be a maintainer: could you drop me an email at m at maxroos period com?

[keewis]

I added a whats-new.rst entry, but I'm not sure if it is too minimal.

If this is okay for now, we could merge?

[keewis]

This should be ready for review / merge once again.

It seems that after the merge of #3706 all that was left to get the top-level function tests to pass was to clean up those tests.

[max-sixty]

Looks good!
(as ever with me & this section of code, without being that confident)

[dcherian]

Yeah, thanks @keewis