The following surpresses a KeyError when a application or a remote
application is missing. Instead we should return None and correctly
handle that a application or a remote-application could be missing.

This should be safe to do so and the fact that an update from the
all watcher should fill in the gaps eventually. If this is not the
case, we should in the future ensure that a remote-application can't
raise index/key errors when data is missing.

The missing data when raised/thrown criples the library from being
useful, instead we should really handle all the cases these exist and
correctly insert None or a fallback.

After having a long internal discussion around this, it's evident that
this will cause errors down the line at some point. Instead we've
decided that it's better to throw a custom error, one that can be caught
individually to force the user of the library to disconnect and
reconnect from the model, to pick up any missing deltas.

This can be done in a way to the examples/model.py where retries can be
forced to happen when an exception is raised.

This code to change this is very rudimental and just exposes a new juju
error - JujuEntityNotFoundError. I've also updated the test to fully
expose the exception at a unit level.

The following changes consume application offers, so that they're
correctly picked up.

This should help identify issues in the future.