MDEV-35397: Fix semisync assertion on slave disconnect/reconnect race

[nadaelsayed11]

Repl_semi_sync_master::remove_slave() clears the entire active-transaction list when the last semisync client disconnects (with rpl_semi_sync_master_wait_no_slave=0), but unlike switch_off() it never bumped rpl_semi_sync_master_off_times. A transaction parked in commit_trx()'s wait loop tolerates a lone disconnect fine, since its own 'no clients left' guard matches remove_slave()'s. But if a slave reconnects before the waiter re-checks, that guard no longer fires, and the waiter finds its entry missing with off_times unchanged, tripping the DBUG_ASSERT that assumes only switch_off() can cause a missing entry. Bump rpl_semi_sync_master_off_times in remove_slave() too, mirroring switch_off(), so the invariant holds and the waiter takes the intended 'skip the wait' exit instead of asserting.