Optimize change handling in tree by invalidating only branches

[mariuswilms]

When a change inside the tree is detected the whole tree is rebuild for the sake of simplicity. Generally we should also be safe to invalidate only the affected node and all its descendants. This will make some parts more difficult, i.e. message will also need to contain the affected node and the descendants need to be calculated (possibly using the flat lookup table). On the other hand gives us better indexing performance until the tree is free to be accessed again (the tree is write locked during rebuilds).