[tonyandrewmeyer]

Adds a private worker that runs one charm's events in a separate Python process, with the event and State crossing as JSON (the encoding from #2828), so that charms whose dependencies conflict can be tested together.

• scenario._isolation.Worker is what the multi-charm model will hold, one per application. It reads the charm's metadata from disk and never imports the charm in the test process. It starts the worker process on the first event and keeps it for later ones, and the unit ID comes with each event, so one process serves every unit.
• A worker that crashes or takes longer than the timeout (60 seconds by default) is killed, raises IsolationError, and isn't restarted. A charm that raises is different: the worker sends the traceback back and stays usable.
• Messages are length-prefixed rather than line-delimited, and the worker moves its responses off file descriptor 1, which it points at stderr, because charms write to fd 1 through subprocesses and C extensions as well as print.
• IsolationError is private for now, in scenario._juju_errors, and feat: add a typed JSON encoding of ops.testing State for worker processes #2828's version error is now a subclass of it. IsolationError moves to testing.errors when the API is exposed.

Some things the proof of concept had are left out: a mode that starts a new process for every event, an idle timer, charm mocking, per-unit filesystems, and telling a charm's own exception apart from a worker failure. Those come in later PRs in the series.

The tests run the worker with fixture charms, including two that need conflicting versions of the same dependency, one that exits the process, one that never returns, and one that writes to fd 1.