Add opt-in parallel setup across nodes

[bgrewell]

Setup steps ran strictly sequentially, so two nodes with slow provisioning
waited on each other for no reason. Each node's chain now runs concurrently
with the others when --parallel-setup is passed: three nodes with a three
second chain each finish in three seconds rather than nine.

Steps for one node stay in order. A node's own steps routinely depend on each
other, so only different nodes overlap.

It is opt-in rather than the default. Sequential order across nodes is the
only ordering mechanism a suite has, so a suite whose step on one node depends
on a step having finished on another is relying on that order without
declaring it. Enabling this by default would break those suites silently,
which is the failure mode this codebase keeps removing.

Two flags force the sequential path and say why: --pause-on-error reads stdin
and cannot prompt for concurrent steps, and --until stops at a named step,
which needs an order to stop in.

Concurrent tasks render without a spinner. The animated completer owns a
terminal line and registers as the process's single active spinner, so two at
once overwrite each other; a quiet completer writes each task's line once, on
completion, under a shared mutex so lines never interleave.

A failure skips the rest of its own node's chain but not the other nodes'.
Every failure is reported and the run exits with the first in node-declaration
order, so the reported error does not depend on scheduling.