WL#1888: Fix SELECT ... LOCK IN SHARED MODE

SELECT .. LOCK IN SHARED MODE
requires that read locks are held after completing the query. This can only
be accomplished by implementing takeOverForRead in NDB API and using this
in the handler.

There is also a dependency to a task where rows that are not used will
call unlock_row so that takeOverForRead is only called when needed.

This means that scan with LM_Read needs to be able to request KEYINFO.

In order not to lock unnecessary rows this work also includes implementing
support for handler::unlock_row
Also, to generalize, an implementation could be done where the current
isolation level decides if locking should be done to support
TRANSACTION_REPEATABLE_READ. Since this would be a performance degradation
the current default isolation level should probably lowered to
TRANSACTION_READ_COMMITTED (as in Oracle) and to issue a warning if
one sets TRANSACTION_REPEATABLE_READ that this impacts performance.