A further reduction in CPU usage by ndb_mgmd with failed data nodes.


ndb/src/common/util/socket_io.cpp:
  Use a nonblocking socket to peek at the buffer to know how much to read for socket_readln.
  
  This saves us a lot of syscalls. By a lot, I've observed about halving the CPU
  usage of ndb_mgmd with 1 failed node on a 6 node cluster.