Issue 880 wait for lost node

[SilinPavel]

This PR provides implementation of #880 It makes some changes to kube configuration, according to discussion in #880 , it also provides implementation of approach of deferred failing of a pipeline that is run on a lost node. Accordingly to this approach:

• New system preference was introduced: cluster.node.lost.tolerant.seconds
• AutoscaleManager will store timestamp when node was lost for the first time
• Each next time it will check that 'timeout' (cluster.node.lost.tolerant.seconds) not exhausted
• If 'timeout' will be exhausted pipeline of this node will be failed
• If lost node will have enough time to recover nothing will happen, and pipeline will continue to work, timestamp of last loss will be reset