Metric needed: what percentage of builds fails on each hypervisor host

This way we'll know if a host is "toast" and it needs to be disabled, otherwise leaving the host online will keep failing whatever builds it receives.

The reason for this ticket is that "ace-fawn" has been spewing I/O errors since more than 10 days ago, apparently failing many of the builds it has hosted, but nobody noticed in order to put the host offline.