The node's last_error disappears briefly on cleaning failure

It is caused by the power off routine, which initialises last_error to None. The field is later restored, but BMO manages to observe and record the wrong value.

This issue is not trivial to reproduce in the product. You need OCPBUGS-2471 to land first, then you need to trigger the cleaning failure several times. I used direct access to Ironic via CLI to abort cleaning (`baremetal node abort <node name>`) during deprovisioning. After a few attempts you can observe the following in the BMH's status:

status:
  errorCount: 2
  errorMessage: 'Cleaning failed: '
  errorType: provisioning error

The empty message after the colon is a sign of this bug.