So, following what chartl said about multiple linear algebra libraries, my thoughts immediately went to intel-mkl conflicting with openblas somewhere…
Using conda list, I noticed I indeed had multiple libraries with either intel, mkl, or openblas in the build or channel name. I decided to uninstall anything containing the former two, and reinstall everything with conda install -c conda-forge ..., so now I have the following, but Pymc3 still gives the same bad energy error:
# Name Version Build Channel
absl-py 0.7.1 py37_0 conda-forge
astor 0.7.1 py_0 conda-forge
attrs 19.1.0 py_0 conda-forge
backcall 0.1.0 py_0 conda-forge
binutils_impl_linux-64 2.31.1 h6176602_1
binutils_linux-64 2.31.1 h6176602_6
blas 2.8 openblas conda-forge
bleach 3.1.0 py_0 conda-forge
bzip2 1.0.6 h14c3975_1002 conda-forge
c-ares 1.15.0 h14c3975_1001 conda-forge
ca-certificates 2019.3.9 hecc5488_0 conda-forge
certifi 2019.3.9 py37_0 conda-forge
cudatoolkit 8.0 3
cudnn 7.1.3 cuda8.0_0
cycler 0.10.0 py_1 conda-forge
dbus 1.13.6 he372182_0 conda-forge
decorator 4.4.0 py_0 conda-forge
defusedxml 0.5.0 py_1 conda-forge
Thinking that some unseen remnant of my intel python distribution was left around, I did a blanket reinstall of anaconda through wget, reinstalled everything into a Python 3.7.3 environment with conda install ... (only being forced to use conda install -c conda-forge theano to get theano 3.6), and it works!
It seems that openblas, either by itself, or when used at the same time as intel-mkl, produces errors on at least Ubuntu 16.04.6.
Thanks for the suggestion, chartl!