Improve shell completion continuous integration tests

[bckohan]

The strategy for testing completion functionality is to run the various shells behind a python psuedo terminal. This works for bash and zshell but does not work for fish or powershell. Right now only install/remove are tested in the CI for fish and powershell. There are also intermittent spurious failures on the bash and zshell tests. There must be a more stable way to test this functionality in an integrated end-to-end way that works for all shells and is deterministic. I would also like to add windows tests.

Until such time as a solution is identified, if changes to shells or the library are made it is best to manually test the fish and powershell completion still works before issuing releases.

This is somewhat ameliorated by shell completion being a nice to have and not a critical feature that prevents using the library.