revisit compatibility with newer generation of pymc3/theano

[bsipocz]

There are various combinations of issues with some of the version combinations: fix compatibility with the latest releases. We need to be more specific of the requirements as the doc currently only states that pymc3 is an optional dependency.

As a workaround, limit and/or pin the version numbers for CI. We used v3.6 to generate the figures for the book, so that could/should be the minimum version, at least for the figure repository.

[bsipocz]

All the book figures using pymc3 are compatible with versions as of 2018 Dec, to be reproduce use, e.g. pypi-timemachine 2018-12-30. However, LinearRegressionwithErrors uses the updated signature of pm.Normal as of v3.7, so the first working timemachine cut-off is 2019-06-01 of pymc3 3.7 and theano 1.0.4

2020-08-01 is OK with pymc3 3.9.2 and theano 1.0.5, however pymc3 3.9.3 is breaking with a BrokenPipeError in multiprocessing with python 3.7&3.8, but passes with 2020-11-15 on python3.9.

pymc3 3.10 also passes on python 3.9 with 2020-12-15, but the very latest 3.11 is breaking with the new theano errors. theano 1.0.11 that is required for pymc3 3.10 is incompatible with numpy 1.12, so that has to be limited in the tox config for now.

(However, to get the tests pass one should get very tight version combinations, I had all sorts of failures locally when I didn't use timemachine to limit everything to a certain date cutoff).

[bsipocz]

python3.10 is not compatible with the arviz version limitation put in place in tox, this issue has to be revisited sooner rather than later.

[bsipocz]

This becomes more critical now that numpy 1.24 is out, and the theano/pymc versions we pin are incompatible with it.