rm_node refactored.

* Now it is possible to remove node with other nodes from this RG offline.
* We remove node from metadata instantly: rm_node is always
  successfull. However, if something went wrong, we ask to run recover() later.
* recover() handles node promotion and ensures that list of subscibed tables is
  up-to-date.

A bunch of python tests fixes included. Now we don't use somewhat weird
testgres's logging service and provide symlinks to node's logs ourselves.

Author: arssher