CI test failing on python 3.8

[dkirkby]

(Copying this over from #192 so we don't forget)

One CI test is failing on python 3.8 but passing on 3.6 and 3.7:

FAIL: test_tan2radec (desimeter.test.test_radec2tan.TestRADEC2TAN)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/circleci/project/py/desimeter/test/test_radec2tan.py", line 139, in test_tan2radec
    assert(np.abs(ddec*3600.)<0.001) # 1 milli-arcsec

[julienguy]

This test is passing on my laptop with python 3.8.3.
dra=-9.853449896013444e-05 arcsec , ddec=-0.00044902475337949 arcsec

[dkirkby]

The failed test log is here. It looks like it uses python 3.8, so updating to 3.8.3 here would likely fix this. Are we ok with abandoning 3.8 now that more recent minor versions don't have this issue?

[sbailey]

Requiring 3.8.3 instead of 3.8.0 is fine, if we know that 3.8.3 has a specific bugfix that is supposed to be fixing the underlying problem (or perhaps more likely, brings in a specific numpy version that fixes the underlying problem). Otherwise if it's just that the problem goes away with 3.8.3 but we don't know why, I'm a bit worried that it could be a more subtle memory alignment etc. issue that we are just failing to trigger with the test on 3.8.3 but haven't really avoided in general.

Please do a little searching and reading python and numpy changelogs to see if you see anything that smells like it is actually a fix to this and not just "it went away".