CI/DOC: pyarrow ImportError when using environment.yml to build the docs

[datapythonista]

In #26648, when we tried to use environment.yml to build the docs in azure instead of ci/deps/travis-36-docs.yaml, we found that pyarrow couldn't be imported, and it failed with the next error :

ImportError: /home/vsts/miniconda3/envs/pandas-dev/lib/python3.7/site-packages/pyarrow/lib.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZNK5arrow6Status8ToStringB5cxx11Ev

There are few differences between the created environments, like Python 3.6 vs 3.7, arrow-cpp being installed from defaults vs conda-forge... But it's unclear what's the exact problem or what is causing it.

Ideally, we'd like to use environment.yml to build the docs, so the environment is exactly the same as the local installations we use.

What we need to do is:

• Identify what is causing the error
• Implement the required changes so pyarrow can be imported
• Replace ci/deps/travis-36-docs.yaml by environment.yml in the doc builds in Travis and Azure.

[rjsee]

Could be inconsistent gcc settings, see
https://issues.apache.org/jira/browse/ARROW-4809
and similarly
ray-project/ray#2483
OTOH might be something else in the environment file upstream that's then leading to this conflict

[jreback]

is this still an issue? @minggli closed #26670 as a patch for this

[datapythonista]

Looks like the problem is not happening anymore. I guess it was a bug on Anaconda's side serving arrow-cpp from the conda-forge channels instead of the defaults conda one.

Checked in the logs of: #26829