During initial start we are careful to start nodes serially, in order, to
ensure the i'th VM's process acquires the i'th node ID. However this makes
startup much slower. We also initialize cluster settings.

When *restarting* the nodes of an existing cluster however, these steps are
not required: the node ID assignments are already persisted in the stores, as
are any previously set cluter settings. Thus we can elect to start all the
processes in parallel and skip init.

Release note: none.
Epic: none.

A common debugging flow is to ask raochprod to run some query on all nodes, eg.
roachprod sql mycluster -- -e 'select version()';

When the invoked statement returns multiple lines it makes sense to
a header above each result set, then the result set as it was returned.

However if the result set is a single line per node, printing just that
line, rather than a roachprod header and the result set's own header,
can make the aggregate result set much easier to read.

Release note: none.
Epic: none.