Fix concurrency and syntax issue in merge-output-to-logs
In a multi-node job, these mv commands run in parallel which makes
the second invocation weird and racy (even though there is a mv -f
there) Put in run_once - we don't need to perform this more than
one time, no matter how many nodes there are in the nodeset.
Also, when there are actually files present, if ! $(ls work_dir/*)
winds up expanding as a command to be run.
When we're operating in the context of a change, we'd like to
upload artifacts and docs to the log server so that people can iterate
on them before publication safely. Add a role which is intended to be
run after fetch-output that moves the content on the executor so that
normal log upload concepts can be applied.