pcp

For the per-process schedstat metrics on Linux, if the running
kernel does not support them we were getting the error handling
wrong - do not close a descriptor we didn't open, and only set
the cluster-fetched flag if that was indeed the case.