Static models marked as not-trained after changing the default predictions processor

Synopsis

When playing with analytics, I realized that the "No teaching" model is not processed by the "predict_models" scheduled task as it should be. Debugging revealed that the model has been marked as not trained in my database - and as such, the task did not pick it. But that should never happen because static model (i.e. those using target based on assumptions, not facts) should be always considered as trained implicitly.

Bug

The problem has been introduced by MDL-59988. When a model is cleared (e.g. when default predictions processor is changed), the model is always marked as not-trained. But that must happen for non-static models only.

Steps to reproduce

1. Check that the "no_teaching" model is correctly marked as trained in the database

   SELECT id,enabled,trained,target FROM mdl_analytics_models;

   It should have both enabled and trained set to 1.

2. Change the default predictions processor from PHP to Python.
3. Check the "trained" value of the "no_teaching" model in the database again.

Expected: The value must be still 1.
Actual: The value has been set to 0.