Remove the AI severity from every artifact on a host, so each reports the severity its sandflies give it again.

The host's own AI rating is withdrawn with it, whether or not any artifact carried an AI severity: until a new verdict is formed the host's severity is the highest severity among its alerting artifacts, and the verdict reports severity_withdrawn. The verdict itself is not deleted and still reports the severity the model gave it.

This clears every AI severity on the host, including one set by a classification requested on an artifact directly -- unlike a new host verdict, which leaves those alone. The un-steered severity is never lost while a verdict is in force, since it is recomputed on every scan, so this restores exactly what the sandflies currently say rather than what they said when the verdict was formed.