Rework harmonic oscillator free energy tests for multistate samplers to run on travis

[jchodera]

Description

This PR reworks the harmonic oscillator free energy tests for multistate samplers, testing them both with and without unsampled states and ensuring they are fast enough to run on travis (fixes #441).

I've also converted the warnings.warn messages about the experimental API to logger.warn so they can be more easily suppressed with the logger level.

[andrrizzi]

This is awesome, thank you!

Just for curiosity, what's the reason why you had to re-generate the cython code? Do you think this might be related to the mixing problems with multiple GPUs we observed (choderalab/yank#1130)?

[jchodera]

Just for curiosity, what's the reason why you had to re-generate the cython code?

That was accidental. It appears that installing the code locally caused that to happen.
Should we exclude that .c file from being included in .gitignore so that it is built at install time?

Do you think this might be related to the mixing problems with multiple GPUs we observed (choderalab/yank#1130)?

Hm. It could be that the view into the arrays somehow skipped every other element. But didn't we only see this issue if we were using MPI? How would MPI be interacting with this?

[jchodera]

The cython changelog doesn't list any bugfixes that seem suspiciously like ours since 0.29.4.

[andrrizzi]

Should we exclude that .c file from being included in .gitignore so that it is built at install time?

I think it's fine to track the .c file. It probably shortens slightly installation and, in principle, it allows us to not rely on the cython dependency.

But didn't we only see this issue if we were using MPI? How would MPI be interacting with this?

Yes, you're probably right. I'm going to take another look at the problem in the next few days though. If this didn't change anything, maybe there was something that changed in mpich that fixed it.