**Problem description**

XA is not handled correctly in combination of GTID_MODE=ON and
--skip-log-bin on the server.

There are two errors whose common reason is a missed out gtid_state's
save() and not performed gtid ownership cleanup.
The server can be in either the master or the slave role that
affects which of execution paths is taken in XA commit and rollback.
The two missed invocations are present in either branch.

**High-level description of fixes**

The missed invocations are added up in the cases when
the XA transaction gets prepared, committed or rolled back.
To remind about the prepared case, the prepared XA is a sort of a
transaction on its own from perspectives of the binary logging and the
gtid_executed table (as instances of the gtid persistency
layer). Therefore while the XA gets merely prepared a statement to
gtid_executed table needs to be fully committed, that is as a separate
transaction.
The patch copes with this task that was actually diagnosed in
bug21452916 report.
In order to separate the gtid record off the prepared XA the patch
designs Attachable_trx_rw a read-write version of THD::Attachable_trx.

**Low-level details** of the changes include:

modified:   sql/handler.cc
  A common through-out sources computation to call gtid_state::save()
  is factored out as a commit_owned_gtids();
  The new function is introduced to ha_prepare() to cover XA PREPARE time
  updating of mysql.gtid_executed.

modified:   sql/rpl_gtid_persist.cc
  Attachable_trx_rw is registered and deactivated (committed).
  XA state is saved, reset for the following execution of
  a statement on 'gtid_executed' table, and restored.

modified:   sql/sql_base.cc
  An assert guarding access to system tables in open_and_lock_tables() is relaxed
  to allow opening specified tables for Attachable_trx_rw

modified:   sql/sql_class.cc
modified:   sql/sql_class.h
  Definition of Attachable_trx_rw and associated interfaces.
  Attachable_trx_rw is defined to base on THD::Attachable_trx.

modified:   sql/xa.cc
  commit_owned_gtids() gtid_state->update_on_*() are deployed in
  few functions to cover all cases of XA COMMIT and ROLLBACK, incl errors,
  one phase.

  A common rule for choosing which of update_on_{commit,rollback}() to
  pick goes as the following:

  Whenever commit_owned_gtids() succeeds that indicates on
  mysql.gtid_executed table successful insertion. And since
  @@global.gtid_executed must be consistent with the table,
  update_on_commit() is chosen in this case.
  Conversely, update_on_rollback() runs only to match a failing
  invocation of commit_owned_gtids().  When at committing or preparing
  there's an RM error (e.g Innodb unilateral failure)
  commit_owned_gtids()is not invoked at all.
  Presence of RM error does not block commit_owned_gtids() invocation
  in the XA-rollback case, because the rollback intent does not
  change.

**Testing**

The patch extends the existing testing base for XA and manual GTID
combination, specifically:

           suite/binlog/t/binlog_gtid_next_xa.test

The above test is augmented with coverage of all affected
commit-rollback execution branches in xa.cc and ha_prepare(), and also
made to run in --skip-log-bin environment.

new:
           suite/rpl/t/rpl_xa_survive_disconnect_lsu_off.test

The slave logics is not directly affected, but still the changes fix
an assert in Diagnostics_area::set_ok_status, when
rpl_xa_survive_disconnect.test is run on slave having
--log-slave-updates=off to face an expected (cos of no resetting of XA
state) error:

  ER_RMFAIL: The command cannot be executed when global transaction is
             in the IDLE state.

The new test is the old one's version requiring the above slave server
option.

**NOTE**: This work does not cover to crash-safety leaving this matter
out for earlier reported BUG#20672719.