When retrieving a SPJ result, which should be returned in sorted order as available
from an ordered index, we need to merge results retrieved from the fragments.

If the sorted index is specified over multiple columns, and the query only select
and/or 'order by' on a leading subset of these columns, the index may still be used
to provide sorted result order for this subset.

The fragment sort-merge used the compare_ndbrecord() function to compare two
records from different fragments. However, this function assumed that all columns
from the sorted index was retrieved and present in the 'result_records'. For any
non-present column it effectively compared non defined data, potentially causing
the reported require to fail in the compare methods.

This patch introduce the usage of the bitmask 'result_mask' which hold information
about which columns being defined in the 'result_record'. compare_ndbrecord() is
enhanced such that it returns when it has compared all the 'valid' columns
from the ordered index preent in the 'result_record'.

Note that compare_ndbrecord() is also used when reading sorted result from
a non-pushed table access, using a sorted NdbIndexScanOperation. However,
for this use case the NdbIndexScanOperation will add any non specified
index column to the specified set of columns to retrieve - effectively
retrieving more columns then first specified. Ww could probably have eliminated
this by using the specified 'result_mask' in this case as well. However, that
would have changed the behaviour of a already public interface (As opposed to
the NdbQuery interface) so we leave that for another bug report.

Patch also remove declaration of compare_ndbrecord() from the public API
header file storage/ndb/include/ndbapi/NdbIndexScanOperation.hpp.
Declaration is already available in the 'friend' declarion NdbReceiver.hpp.

Reviewed by: Frazer Clement <frazer.clement@oracle.com>

(cherry picked from commit 49b3eaea2bf1901cbe7c774fd9553c4a22dc061d)