Bug#29525670 NDB.NDB_READ_BALANCE FAILS 96 SECONDS AFTER NDB.NDB_MULTI

Testcase ndb_read_balance aims to capture some of the behaviour of
the different table distribution types (READ_PRIMARY, READ_BACKUP,
FULLY_REPLICATED) in test output.

However it was noted that it was not stable.

Some reasons for this :
 - Read Backup + Fully Replicated tables may read a primary or
   backup fragment replica depending on :
   - The lock mode (long-locking reads must go to the primary)
   - Whether the chosen TC is local to the primary replica or not
 - Unhinted transactions are non-deterministic other than that
   over some number of iterations they should exhibit round-robin
   behaviour over relatively 'local' data nodes
 - Hinted transactions for read-backup and fully replicated tables
   are also liable to read Primary or Backup nodes depending on
   locality.

The test is modified to :
 - Show that for READ PRIMARY tables, the primary replicas are
   effectively always chosen
 - Show that for READ BACKUP + FULLY REPLICATED tables, the primary
   replicas are always chosen for locking reads
   - For non locking reads, a balance of PRIMARY + BACKUP replicas
     are chosen.

This modification should remove the test dependencies on preceding
transactions, endinanness etc.

Change-Id: I5f6e41ae888a0279f0b9b316e214e8643996410f

Author: frazerclement