The BKA implementation in mysql-5.6 has introduce the new
member variable 'mrr_length_per_rec' in 'class ha_statistics'.
The BKA implementation use it to predict whether the MRR buffer
used by the BKA algorithm will become full if we batch another row.

This variable was assumed to be initialized by handler::info()
but this change was not implemented in ha_ndbcluster::info().

Furthermore, the variable was not initialized by the
ha_statistics C'tor so it contained garbage causing
unpredictable behaviour of the 'BKA full' detection.

This could cause the BKA algorithm to incorrectly detecting
the 'full' condition after only a single row had been batched.
Effectively disabling the entire batch logic and causing 
excessively extra Mysql <-> NDB roundtrips.