bats: treat run as setting status, output and lines

[kolyshkin]

In bats, $status, $output and $lines (and, with --separate-stderr,
$stderr and $stderr_lines) are set by the run helper. ShellCheck,
however, only knows that a @test somehow sets them. This does not help
a helper function in a separate (non-.bats) file which calls run and
checks the results, so it gets SC2154. On the other hand, $stderr and
$stderr_lines are treated as always set, even in non-bats scripts.

This PR:

• teaches ShellCheck that run sets these variables (stderr and
  stderr_lines only with --separate-stderr);
• stops treating $stderr and $stderr_lines as always set;
• does not report variables set by run (or by a @test, see below)
  as unused (SC2034), as checking them is up to the caller;
• removes the pretend references to these variables from a @test.
  They were only needed to avoid SC2034 for the above, but they also
  hid SC2034 for manual assignments which are never used (like
  output=$(cmd) instead of run cmd, or a leftover local status),
  and caused SC2031 to be reported on a @test line.

A @test still counts as setting these variables for SC2154. I tried
removing that, but on ~1000 .bats files from various projects
(podman, buildah, cri-o, bats-core, etc.) it resulted in ~730 new
SC2154 warnings, as tests commonly use wrappers around run (like
run_podman) defined in files loaded via bats' load, which ShellCheck
does not follow.

Testing

Compared with the current master on the same ~1000 .bats files:

• 198 SC2030/SC2031 and 2 SC2154 warnings are gone (the ones I checked
  were all false positives, e.g. reading $output after a fresh run);
• 1 new SC2030/SC2031 pair, which is a real bug:
  ( cd dir && run git reset ... ); [ "$status" -eq 0 ];
• 2 new SC2034 (local status where status is never read, and a bats
  test fixture which deliberately sets status in teardown).

In runc integration tests, this makes two # shellcheck disable=SC2154
annotations (in a helper calling run) unnecessary. It also found a few
pointless assignments (output=$(cmd) instead of run cmd, and unused
local status), which are being fixed on the runc side.

Note this touches the same test area as #3550, so there will be a
trivial conflict (adjacent prop_ lines) when merging both.

Refs #356

[kolyshkin]

@e-kwsm @koalaman PTAL; this improves checking bats files.