Fixes DINF-82; This was very much a rabbithole. A few things:

• The race that @bobheadxi mentioned here https://github.com/sourcegraph/sourcegraph/pull/63405#discussion_r1648180713 wasn't from *output.Output being unsafe, but outputtest.Buffer as it happened again (see DINF-82)
• Fixing the above led to discover 👇
• There something messed up with cmds.start(), which sometimes ends up printing the command output after the exit message instead of before.
  • The crude sort.Strings(want|have) that was there already fixes that.
  • Dug around for a while, but didn't find anything yet, because that duct tape is probably hiding something fishy.
• And without the sleep, it's possible to read the output from the outputtest.Buffer before the command outputs get written to it.
  • The time.Sleep(300 * time.Milliseconds) mitigates/hides that problem.

At least, this shouldn't blow up in CI and buys us time to fix the whole thing. We're tracking this in DINF-104. And out of 200 runs, I also stumbled once, on a race in progress_tty, tracked in DINF-105 (that packages is originally meant to be used by src-cli and was re-used for sg 3 years ago).

I'm pretty unhappy about the solution, but a bandage is better than nothing. While ideally, we should really reconsider dropping std.Output entirely in sg and use the good stuff from github.com/charmbracelet instead because we don't want to spend too much time on arcane terminal things ourselves, I'm much more about concerned the concurrency issues mentioned above. We'll circle back on this.

Test plan

CI + sg bazel test //dev/sg:sg_test --runs_per_test=100