Mapping need to be performed when reading data using indexes or primary
key whose column order is different than the column order of the NDB
table. For unique indexes an attrid map is built when the table is
opened which is then used for each subsequent read. For primary key
reads the attrid map is built for every read.

Since a way to cache the attrid map already exists, it seems like a good
idea to build also the attrid map for primary key when opening the
table. This way the implementation will have a more uniform code path
for reads. Since the attridmap for indexes which have the same column
order are well known, this also brings an opportunity to implement an
optimization to not build an attrid map for the most common case.

This table shows indexes whose columns are in different order than the
NDB table column order:
  CREATE TABLE t1 (
    a int, b int, c int, d int, e int, f int, g int, h int,
    PRIMARY KEY(d,b,c),
    UNIQUE_KEY(e,d,c)
  ) engine = ndb;

Fix by building cached attrid map also for primary key when opening the
table and use it for subsequent reads.

Also add the optimization to avoid building the attrid map when the
column order are identical, this reduces the amount of memory which need
to be allocated for the common case where columns are in same order.

Change-Id: I9be76fc33a83ff82a54613decc0af8efe976f42f