We probably shouldn't set it to None before checking whether it's None,
that seems less efficient than it could be.

On the run of a derived test class, cls._DBusTestCase__datadir calls up
into the parent class for the datadir and removes it. The
subsequent statement cls._DBusTestCase__datadir = None sets it on the
derived class, not on DBusTestCase which remains at the original value.

Doesn't matter too much since the next derived test class will repeat
that process but we leave the parent datadir set despite having removed
it which seems like a bad idea.

See the comment in get_services_dir(), the datadir is always on
the DBusTestCase class, so let's mirror that behavior here.

A slightly weird side-effect here: start_dbus was called on the
DBusTestCase class - thus the session/system bus pid was set on that
class.

But tearDownClass() was called with the derived class from the test, so
while the getters for variables resolved correctly,
   setattr(cls, f"{bus_type}_bus_pid", None)
would set TestFoo.session_bus_pid = None and leave
DBusTestCase.session_bus_pid at whatever value it had before.

Fix this and add an assert to ensure we never start a session/system bus
twice in the same test class.

No real functional changes, this prepares the way for some extra asserts
in the future and the behavior matches that of unittest-style classes
which are always subclasses of DBusTestCase.

This is a potential break but it's unlikely to affect any old-style
unittest classes since they would be subclasses by default. It may
affect new-style pytest classes, if any exist, but OTOH it
can catch confusing errors like setting a variable on the DBusTestCase
class but then unsetting it on the derived class during a test.

This is what they are anyway, so let's do this and avoid the possibly
accidental wrong class handling fixed a few commits ago.

This patch introduces the BusType class which handles the system/session
buss differences including whatever is independent of who started the
actual daemon such as getting a connection.

It moves spawn_server and spawn_server_template to standalone functions
since those do not have any dependency on anything else in
DBusTestCase.

It adds the PrivateDBus class which wraps a single instance of a DBus
daemon, with a context manager so we can do things like:

    with PrivateDBus(BusType.SESSION) as bus:
        ....

and have it self clean up automatically. Unlike the previous approach,
we don't just fork in the background and kill via PID, we keep a
reference to the process and kill it via Python's API.

Finally, it changes the DBusTestCase class to use a single instance of
those objects as its DBus server when created with start_session_bus()
or start_system_bus(). Most of that class is now a wrapper around the
PrivateDBus object (or the BusType object where appropriate).

PrivateDBus also wraps some functions to make it look more like a
DBusTestCase object to keep compatibility with the current pytest
behavior.

With all that, we can now return a PrivateDBus object from the pytest
fixtures and have better insulation than before where the DBusTestCase
class was responsible for singleton handling.

Let's move spawn_server and spawn_server_template into an object with a
context manager, greatly simplifying them where fixtures are being used.
Note that both are renamed into:
 SpawnedMock.spawn_for_name()
 SpawedMock.spawn_with_template()

Some of these could've only been exposed by the recently added pytest
fixtures - those have never been in a release so let's drop them.

deciseconds is a rather unusual unit. Let's switch the real API we now
have to use seconds: float like time.sleep() and wrap our old API around
that.