Add a metric for tracking erroneous vCPU kicks

[ludfjig]

If a signal is delivered very late as a result of calling kill(), it can interrupt a guest call that was not intended to be cancelled. This results in the vcpu getting "kicked", but then re-enters the guest VM immediately because the new call was not intended to be cancelled. It would be good to track how much this happens.

See #1024 (comment) for original comment