drop python 3.5

[aloctavodia]

Do not test Python 3.5. ArviZ will also drop 3.5 I think that was the main reason to include 3.5 in PyMC3 tests.

[codecov]

Codecov Report

Merging #3700 into master will decrease coverage by <.01%.
The diff coverage is n/a.

@@            Coverage Diff            @@
##           master   #3700      +/-   ##
=========================================
- Coverage   89.91%   89.9%   -0.01%     
=========================================
  Files         134     134              
  Lines       20269   20269              
=========================================
- Hits        18224   18223       -1     
- Misses       2045    2046       +1

Impacted Files	Coverage Δ
pymc3/distributions/distribution.py	94.95% <0%> (-0.3%)	⬇️

[ColCarroll]

I'm not terribly against dropping support for the next release (definitely should include a line in RELEASE-NOTES), but for reference, here is the Python support schedule for EOL:

https://devguide.python.org/#status-of-python-branches

Python 3.5: 13 September, 2020
Python 3.6: 23 December, 2021
Python 3.7: 27 June, 2023
Python 3.8: October 2024

[aloctavodia]

I was following NumPy schedule. We agreed to follow it in ArviZ, but maybe we are very anxious people :-). They will drop 3.5 in their next release (Jan 07, 2020). https://numpy.org/neps/nep-0029-deprecation_policy.html

[aloctavodia]

Also because of this #3693. But I am OK with the current progress bar, so no big deal

[rpgoldman]

Totally down with this, because it would mean I won't have to use the bad old syntax for variable type hints, and can move on to the more streamlined modern syntax...

[ColCarroll]

I love this numpy table, and that they have done the hard thinking for us! If I have time, I will try to find a good space in the docs to link to this. I guess the policy should be "any releases after a date in here will have these minimum requirements."?

This also suggests that we only support numpy>=1.15. For reference:

pymc3 >=1.13.0
arviz >=1.12
theano >=1.9.1
pystan >=1.7

this would probably help with testing and with tooling to start restricting versions a bit more aggressively.

[twiecki]

Also in favor of following numpy and dropping 3.5.

[ColCarroll]

I think we should merge this, but also should enable testing on 3.7 and 3.8 (which I just found out has been available for 3 weeks)