Add tests for interactive `bst init` command using
[pexpect](https://pexpect.readthedocs.io).

This also involves a number of changes to tests and other parts of the
codebase since they were hacking about wit API that shouldn't have
existed.

This updates all dependencies on the project, which is mainly needed
by python3.8 but can be done independentely.

This also disables multiple false positive lint errors and disable a
new check that we don't need.

This gives a potentially more explicit understanding of what went
wrong, and pytest can give better information about that exception
than just us asserting the return code.

As discussed over the mailing list, reformat code using Black. This is a
one-off change to reformat all our codebase. Moving forward, we
shouldn't expect such blanket reformats. Rather, we expect each change
to already comply with the Black formatting style.

Currently, the `BaseRemote` would call `_check()` on the children, which
states that errors should be sent back as a string. However,
`BaseRemote` doesn't check the return of `_check()`.

This changes the contract so that subclasses throw `RemoteError`
themselves.

This also fixes the `ArtifactShare` and add a test.

Use the documented path [1] to `pytest_cov.embed.cleanup_on_sigterm()`,
to avoid crashing on some versions.

It turns out that pytest_cov v2.6.1 on my machine doesn't like the way
that we were accessing cleanup_on_sigterm(). Access it in such a way
that we will either get the function or an ImportError, as per the
documentation.

[1]: https://pytest-cov.readthedocs.io/en/latest/subprocess-support.html

Remove a couple of cases where it's possible to make the main test
process hang, waiting for something to appear on a queue.

Raise a friendlier exception, earlier, if there was a problem starting
the server process.

Note that unittest.MagicMock is unfortunately not pickable, so if one
tries to cross over to a ChildJob then we're in trouble.

Sacrifice some convenience and a bit of safety by implementing our own
_DummyTask instead of using MagicMock.

Don't try to pickle the 'simple_task' context manager, if it's been
overridden when creating a test context.

To ensure that we only disable element loading task progress reporting
for very specific code paths, we need to teach the test suite to be a
bit smarter.

For this reason we now mock a _Task object and return it in our mock
context's relevant method invocations.

Other code paths that deliberately invoke the loader without task
reporting now mark their loads with NO_PROGRESS.