The bug was in JOIN::join_free which was wrongly determining that
all joins have been already executed and therefore all used tables
can be closed.


mysql-test/r/subselect_innodb.result:
  - test results fixed (Bug#12736 "Server crash during a select
mysql-test/t/subselect_innodb.test:
  - a test case for Bug#12736 "Server crash during a select": test
  that ha_index_or_rnd_end and mysql_unlock_tables are called
  for all used tables in proper order.
sql/item_subselect.cc:
  - implement subselect_union_engine::is_executed
sql/item_subselect.h:
  - implement Item_subselect::is_evaluated. This function is used
  to check whether we can clean up a non-correlated join of a subquery
  when cleaning up the join of the outer query
sql/sql_lex.h:
  - declare st_select_lex::cleanup_all_joins
sql/sql_select.cc:
  - remove an argument from JOIN::join_free, it's now not used
  - reimplement JOIN::join_free to not unlock tables if there
    is a subquery that has not yet been evaluated. Make sure that the
    new implementation calls ha_index_or_rnd_end for every table in
    the join and inner joins, because all table cursors must be closed
    before mysql_unlock_tables.
sql/sql_select.h:
  - JOIN::join_free signature changed
sql/sql_union.cc:
  - implement a helper method st_select_lex::cleanup_all_joins, which
    recursively walks over a tree of joins and calls cleanup() for
    each join.