Probably fixed flaky tests waiting on collect()

this is a sus fix, and it's still possible for new tasks to get added between the task crashing and the worker getting flagged as crashed.
however, this ensures that the worker's state will be consistent with the tasks collected by collect() so that collect() can properly detect errors.
this theoretically fixes workers-no-submit-after-task-crash.phpt flaking under valgrind.

src/routine.c

@@ -121,12 +121,17 @@ static void* pmmpthread_routine(pmmpthread_routine_arg_t* routine) {
 					zval connection;
 					pmmpthread_object_connect(PMMPTHREAD_FETCH_FROM(Z_OBJ(original)), &connection);
 					zend_result task_result = pmmpthread_routine_run_function(PMMPTHREAD_FETCH_FROM(Z_OBJ(connection)));
+					if (task_result == FAILURE) {
+						//this needs to happen before we put the crashed task on the GC queue
+						//so that waiting on collect() can reliably detect errors
+						//this ensures that tests like workers-no-submit-after-task-crash.phpt work correctly
+						pmmpthread_monitor_add(&ts_obj->monitor, PMMPTHREAD_MONITOR_ERROR);
+					}
 					pmmpthread_worker_add_garbage(thread->worker_data, &done_tasks_cache, &connection);
 					zval_ptr_dtor(&connection);
 
 					if (task_result == FAILURE) {
 						//we may have run out of memory or some error that left the interpreter in an unusable state
-						pmmpthread_monitor_add(&ts_obj->monitor, PMMPTHREAD_MONITOR_ERROR);
 						break;
 					}
 				}