support for py3.8 and np1.18

[sbailey]

This PR cleans up some warnings from python 3.8 + numpy 1.18.5:

• numpy is getting more picky about not using where on scalars, thus necessitating
  more verbose if np.isscalar(x) ... else ... code.
• avoid calling np.median on empty slices (previously that would return a NaN, so
  still using that if the slice was empty).

Tests pass at NERSC using both the current desi environment with python/3.6 + numpy/1.16.4 and a test python/3.8 + numpy/1.18.5 environment, but I'm not sure if the tests cover the Gaia NaN and QA empty slice cases so please take a look.

[geordie666]

I ran some "typical" target/cmx cuts and QA cases in this branch and didn't trigger any errors. It's always possible, of course, that there are some cases we don't "typically" encounter, or some malicious cases I didn't think to run.

If we encounter some fragility in the future, we should fix the code and add a unit test to catch the specific failure mode. But, for now, I'm happy to approve this as-is.

@sbailey: I wasn't sure if you wanted to leave the:

warning.filterwarnings('error', '.*Calling nonzero on 0d arrays*')

set-ups in test_cmx.py and test_cuts.py. I'd probably remove them, as they seem specific to your test cases?

[sbailey]

Thanks @geordie666 . In this case I do want to leave the "warnings as errors" for the tests, since I think if we accidentally do re-introduce this warning we want to fix it since it is triggered by a deprecated "feature" that will change in the future. We can turn it back off if we hit a legitimate case where that "nonzero on 0d arrays" warning is triggered that we don't want to work around (I'm not sure what that case would be...)