have a second look at 'mean', 'l2norm' and 'interpolate'

'mean' currently produces wrong results. A first test showed that for an unstructured 3d test grid the integration volumes did not sum up to one (as it would be expected). 2d tests didnt show this behaviour (yet). As the syntax of 'mean' is so simple, the bug is likely to be found somewhere else. Due to the similar structure of 'l2norm', we should test this function as well