Draft patch that fixes and a sketches test cases for:

Bug#20837 Apparent change of isolation level during transaction,
Bug#46527 COMMIT AND CHAIN RELEASE does not make sense,
Bug#53343 completion_type=1, COMMIT/ROLLBACK AND CHAIN don't 
preserve the isolation level
Bug#53346 completion_type has strange effect in a stored 
procedure/prepared statement

Make thd->tx_isolation mean strictly "current transaction 
isolation level"
Make thd->variables.tx_isolation mean "current session isolation
level".
The current transaction isolation level is now established
at transaction start. If there was a SET TRANSACTION
ISOLATION LEVEL statement, the value is taken from it.
Otherwise, the session value is used.
A change in a session value, made while a transaction is active,
whereas still allowed, no longer has any effect on the
current transaction isolation level. This is an incompatible
change.
A change in a session isolation level, made while there is
no active transaction, overrides SET TRANSACTION statement,
if there was any.
Changed the impelmentation to not look at @@session.completion_type
in the parser, and thus fixed Bug#53346.
Changed the parser to not allow AND NO CHAIN RELEASE,
and thus fixed Bug#46527.
Changed the transaction API to take the current transaction
isolation level into account:
- BEGIN/COMMIT now do preserve the current transaction
isolation level if chaining is on.
- implicit commit, XA COMMIT or XA ROLLBACK or autocommit don't.

sql/sql_priv.h

@@ -215,6 +215,11 @@ enum enum_var_type
 
 class sys_var;
 
+enum enum_yes_no_unknown
+{
+  TVL_YES, TVL_NO, TVL_UNKNOWN
+};
+
 #ifdef MYSQL_SERVER
 
 #endif /* MYSQL_SERVER */