The NdbTransaction::getNdbErrorOperation() method returns a pointer to the
NdbOperation object which has most recently had an error condition, if
any.

MySQLD uses the internal NdbTransaction::releaseCompletedOperations() method
to release NdbApi side operation objects after a batch is executed.

In cases where NdbTransaction::getNdbErrorOperation() is called after
the operation with an error has been released by
NdbTransaction::releaseCompletedOperations(), it is possible for
invalid data to be read.

This problem is corrected by invalidating the NdbTransaction's
error operation pointer when releasing completed operations.

Tested by :
 - Adding debug-only extra invalidation of operation members
   when operations are released
 - Using Insert Ignore and TransacionInactiveTimeout to :
   - Cause an operation error without aborting the transaction
   - Hit a non-operation related error on transaction COMMIT