BUG#13987 Cluster: Loss of data nodes can cause high CPU usage from ndb_mgmd

smaller patch for 5.0.

complete patch going to 5.1 due to more intrusiveness for 'list sessions' etc


ndb/include/mgmapi/mgmapi.h:
  add internal get_fd to use in test
ndb/include/util/InputStream.hpp:
  - add this weird startover member to SocketInputStream
      - this helps work out if we've read a newline yet and should start inserting
           into buffer from the start
ndb/src/common/util/InputStream.cpp:
  remove evil, add more.
  
  keep track internally we've retrieved a newline yet (m_startover)
ndb/src/common/util/Parser.cpp:
  change way detecting of NoLine
  
  remove some trailing whitespace that was uglying the place up a bit
ndb/src/common/util/socket_io.cpp:
  Always retrieve data from the OS so that we instantly get EOF on disconnect
  and don't end up spinning looking for a newline.
ndb/src/mgmapi/mgmapi.cpp:
  add internal ndb_mgm_get_fd() for internal testing
  
  handle 'node status' a bit better
ndb/test/ndbapi/testMgm.cpp:
  Add test for MgmApiSession disconnection (mgmd at 100%)
  
  not fully automated due to smaller patch for 5.0
  
  will be complete in 5.1

ndb/include/util/InputStream.hpp

@@ -40,6 +40,7 @@ extern FileInputStream Stdin;
 class SocketInputStream : public InputStream {
   NDB_SOCKET_TYPE m_socket;
   unsigned m_timeout;
+  bool m_startover;
 public:
   SocketInputStream(NDB_SOCKET_TYPE socket, unsigned readTimeout = 1000);
   char* gets(char * buf, int bufLen);