fix: Correctly handle recursive filtering

[mwcampbell]

I discovered that this was a problem when testing a GTK app with the actual AT-SPI adapter. Attempting to unregister the interfaces for a removed node resulted in a D-Bus error saying that the Accessible interface hadn't been registered in the first place. The problem was that the node wasn't registered in the first place, since its parent had the hidden flag set, meaning that the parent's filter result was ExcludeSubtree. But when the node was removed, its filter result was Include, since calling filter on the node didn't check the parent, and the node itself didn't have the hidden flag set.

So now, common_filter checks the parent.

I added add_subtree and remove_subtree because a callback to node_updated on a single node can indicate a change in the visibility of the whole subtree. And I added the added_nodes and removed_nodes sets to ensure that no matter what sequence of events happens, we don't try to register or unregister a node twice in a single set of changes.

[DataTriny]

This is an edge case I admitedly didn't think at the time. Thanks.