(as part of the auto_increment cleanup of WL#3146; let's not be
sad, that monster push still removed serious bugs):
one problem with INSERT DELAYED (unexpected interval releases),
one with stored functions (wrong auto_inc binlogging).
These bugs were not released.


mysql-test/extra/binlog_tests/binlog_insert_delayed.test:
      more tests of binlogging of INSERT DELAYED: with multi-row INSERTs.
      I identified why sleeps are needed to get a repeatable row-based
      binlogged: because without sleeps rows sometimes get groupped
      and so generate different row based events.
mysql-test/extra/rpl_tests/rpl_foreign_key.test:
  don't forget to drop tables on slave too, otherwise it leaves
  an orphan innodb table leading to rpl_insert_id failing sometimes
  (like in pushbuild "sapsrv2 -max").
mysql-test/extra/rpl_tests/rpl_insert_id.test:
      testing that if some statement does not update any row, it does
      not pollute the auto_inc binlog variables of the next statement;
      the test has to use stored procedures because with plain statements,
      mysql_reset_thd_for_next_command() does the resetting (and thus
      there is no problem); mysql_reset_thd_for_next_command() is not
      called inside routines.
mysql-test/r/binlog_row_binlog.result:
  result additions
mysql-test/r/binlog_statement_insert_delayed.result:
  result additions
mysql-test/r/binlog_stm_binlog.result:
  result additions
mysql-test/r/rpl_insert_id.result:
  result additions
mysql-test/r/rpl_loaddata.result:
      With the change to log.cc reverted, the result changes and is better:
      the change to log.cc had caused some INSERT_ID events to disappear
      though they were necessary (but testsuite could not catch that because
      it's single-threaded).
mysql-test/r/rpl_ndb_insert_ignore.result:
  NDB is now like other engines regarding INSERT IGNORE: autoincrement
  values which caused a duplicate key are re-used for next row, not lost.
  rpl_ndb_insert_ignore.result is now identical to rpl_insert_ignore.result.
sql/log.cc:
      LOAD DATA INFILE is binlogged as several events, and the last of them must
      have the auto_inc id. So it's wrong to reset the auto_inc id after every
      binlog write (because then it's lost after the first event of LOAD
      DATA INFILE and so missing for the last one)/
      Another problem: MYSQL_LOG::write() is not always called (for example
      if no row was updated), so we were missing reset in some cases.
sql/sp_head.cc:
      SELECT func1(),func2() generates two binlog events, so needs to
      clear auto_increment binlog variables after each binlog event
      (it would be more natural to clear them in the log write code,
      but LOAD DATA INFILE would suffer from this see the cset comment
      for log.cc). Without the clearing, the problem is:
      > exec func1()
      >> call cleanup_after_query() (which does not clear our vars here)
      >> binlog SELECT func1()
      <
      > exec func2()
      and so SELECT func2() is binlogged with the auto_inc of SELECT func1().
sql/sql_class.cc:
      after every statement we should clear auto_inc variables used for
      binlogging, except if this was a function/trigger (in which case
      it may be "INSERT SELECT func()", where the cleanup_after_query()
      executed in func() should not reset the auto_inc binlog variables
      as they'll be necessary when binlogging the INSERT SELECT later).
sql/sql_insert.cc:
      - as INSERT DELAYED uses the same TABLE object as the delayed_insert
      system thread, we should not call ha_release_auto_increment()
      from INSERT DELAYED (and btw it's logical as we reserve nothing
      as we don't perform the insert). Calling the function caused us to
      release values being used by the delayed_insert thread.
      So I do the call only if this is a non-DELAYED INSERT.
      - Assuming two INSERT DELAYED which get grouped by the delayed_insert
      thread, the second may use values reserved by the first, which is ok
      per se, but is a problem in statement-based binlogging:
      the 2nd INSERT gets binlogged with the "interval start" value
      of the first INSERT (=> duplicate error in slave).
      - no reason to ha_release_auto_increment() after every inserted row
      in INSERT SELECT; more efficient to do it only when the statement ends
sql/sql_parse.cc:
  a comment