Problem :
---------
If "ALTER TABLE ADD UNIQUE index" operation fails, the command adds
the new index in dict_table_t in "Prepare" stage and attempts to
remove the index in "Rollback" stage. We prevent the removal if any
concurrent operation is going on the table. This is done by checking
if dict_table_t is referenced by other objects (n_ref_count > 1).

In 5.7 partitioned table, we keep the dict_table_t objects for the
partitions in shared place and for multiple TABLE objects, referred by
concurrent operations, we keep only one reference for dict_table_t
object (n_ref_count = 1).

It causes the ALTER TABLE to remove the index entry and concurrent
INSERT hits the ASSERT as ins_node_t::entry_list count doesn't match
the number of indexes in dict_table_t.

Solution :
----------
Increment already cached dict_table_t reference count for all the
partitions during ha_innopart handler initialization.

a. New dictionary interface to increase and decrease reference count
   for already cached dict_cache_t element.
      dict_table_t::acquire
      dict_table_t::release
b. Acquire [dict_sys->mutex] during ope/close ha_innopart handler and
   increase/decrease dict_table_t reference count for all partitions.
      Ha_innopart_share::open_table_parts
      Ha_innopart_share::close_table_part

Reviewed-by: Marko Makela <marko.makela@oracle.com>
Reviewed-by: Mattias Jonsson <mattias.jonsson@oracle.com>

RB: 8463