Don't cleanup the datadir's on test failure.
I hope that this should make it easier to debug rare-failing tests, as we'll at least have the debug.log's available to go through.
@MarcoFalke: After this, we could perhaps add to the rpc-tests.py script some functionality to tail the last few hundred lines of debug.log for each node if a test fails, so that if a job fails in travis, we'll have more info to look at?