Watcher dies with uncatchable ConnectionClosed error

[johnsca]

Originally reported via conjure-up's Sentry:

ConnectionClosed: WebSocket connection is closed: code = 1006, no reason.
  File "juju/model.py", line 648, in _start_watch
    self.loop)
  File "juju/utils.py", line 94, in run_with_interrupt
    result = [f.result() for f in done if f is not event_task]
  File "juju/utils.py", line 94, in <listcomp>
    result = [f.result() for f in done if f is not event_task]
  File "juju/client/facade.py", line 392, in wrapper
    reply = await f(*args, **kwargs)
  File "juju/client/_client1.py", line 59, in Next
    reply = await self.rpc(msg)
  File "juju/client/overrides.py", line 104, in rpc
    result = await self.connection.rpc(msg, encoder=TypeEncoder)
  File "juju/client/connection.py", line 238, in rpc
    result = await self.recv(msg['request-id'])
  File "juju/client/connection.py", line 175, in recv
    return await self.messages.get(request_id)
  File "juju/utils.py", line 63, in get
    raise value
  File "juju/client/connection.py", line 183, in receiver
    loop=self.loop)
  File "juju/utils.py", line 94, in run_with_interrupt
    result = [f.result() for f in done if f is not event_task]
  File "juju/utils.py", line 94, in <listcomp>
    result = [f.result() for f in done if f is not event_task]
  File "websockets/protocol.py", line 299, in recv
    raise ConnectionClosed(self.close_code, self.close_reason)

This should be reported to the Monitor. However, we probably also need to have block_until surface closed connections instead of blocking indefinitely.