CI: include test against nightly numpy

[martinfleis]

Closes #1937

[martinfleis]

Once everyone is happy with the approach I took here, I'll mirror it to branca.

[ocefpaf]

I'm OK with this approach but I'm afraid our tests are not robust enough to catch any failures. The way we use numpy here is unchecked and will produce passing status even when it is wrong. As you confirmed in python-visualization/branca#163.

[Conengmo]

Agree with Filipe that the code tests probably don't help here. But it doesn't hurt to test against the nightly releases either?

Not sure if I would run the tests a second time. I'd think that makes the output harder to understand. Or is this a temporary addition? In that case it doesn't matter as much.

We might by the way have more success with the notebook/docs selenium tests, as those also check for Javascript errors (or at least I hope that's the case)

[martinfleis]

Not sure if I would run the tests a second time. I'd think that makes the output harder to understand. Or is this a temporary addition? In that case it doesn't matter as much.

Not sure what your suggestion here is then. Split it into two separate runs? Generally, it is better to test against nightly versions of our dependencies, assuming the test will catch issues, so I would try to keep it as permanent. In other projects I maintain, we have dev as another environment alongside the others (latest, oldest...)

[ocefpaf]

@martinfleis I can add this later but, if you want to tackle this now, let's add an experimental label to that matrix item to make it easier to find in the logs. I did something like that in https://github.com/Unidata/netcdf4-python/pull/1319/files#diff-83620a0d7a95a32e599b5b5b89eabf1419cb0e3ffb5212d15f87f5bd3a452730R17

[Conengmo]

Looks good, let's do it!