Add support for latest versions of h5py, sympy, numpy, and pytest

[ngoldbaum]

This fixes a number of issues caused by updates to upstream libraries we depend on. This is a bit of a grab-bag PR that makes the tests pass again on the latest versions of our dependencies. Unfortunately I can't easily issue these changes as separate PRs since the tests will be broken without all of these coming in at once. If you have questions about reviewing this I'm happy to help out.

To catch issues caused by upstream changes more quickly in the future I've updated the tox config to always recreate a virtualenv. This will ensure we get test failures on the cron job test whenever an upstream release breaks our test suite.

I also made the following updates to our tox config:

• Add testing for python 3.8, update older python minor versions
• Remove pytest version constraint now that issue with pytest-doctest-plus has been fixed.
• Pin coverage to be older than coverage 5.0, which allows us to work around Coverage in parallel failing nedbat/coveragepy#883

There also some code changes, in particular

• Avoid deprecation warnings from h5py by always specifying the file mode in h5py.open.
• Fix usage of rationalize transformer in our usage of sympy.parse_expr. Before we were not using it correctly since we needed to run auto_number first.
• Fix latex output for non-float unit scalar multiples.
• Add support for NumPy 1.8 by fixing handling of the out keyword argument for ufuncs that have multiple outputs.
• Fix warnings from Pint caused by our implicity using Pint.quantity.__array__.
• Make the unit registry saved to HDF5 files smaller to avoid the attribute size limit triggered in newer versions of h5py. Fixes h5py throwing errors on yt-4.0 branch yt#2404.

[jzuhone]

Add support for NumPy 1.8 by fixing handling of the out keyword argument for ufuncs that have multiple outputs.

Do you mean NumPy 1.18?

[ngoldbaum]

Yes, sorry for the typo.

[jzuhone]

@ngoldbaum do you want anyone else to have a look at this before merging?

[ngoldbaum]

I was a little worried about the test failures in yt from the changes in this PR. I'm not convinced that the yt failures are something that can't happen with unyt, since unyt's Unit class doesn't subclass sympy's Expr class. I'll go ahead and merge this in now. Thanks for taking a look John.