Add skip_test and --no-skip; fix a suite that reported green while dying

[helly25]

Add skip_test and --no-skip; fix a suite that reported green while dying

skip_test ends the test it is called from. Callers previously wrote
probe_capability && return at every site, because a bash function can
only return from itself - a skip helper that merely returns leaves the rest
of the body running. Test bodies now run in a SUBSHELL whose exit code
carries the verdict out (77 = skipped), which is what makes an immediate
abandon possible. Expectation errors live in a variable the subshell cannot
write back, so they are folded into that exit code inside it.

--no-skip turns every skip into a failure, for an environment that
GUARANTEES the capability a test probes for. That is not hypothetical: a
FUSE test suite skipped every mount and reported "PASSED in 0.1s" on a
machine that could mount, hiding the whole kernel path.

Adding the subshell surfaced something worse, which this also fixes. THIS
REPOSITORY'S OWN TEST SUITE WAS DYING AT CASE 5 OF 12 AND REPORTING PASS.
Two independent bugs stacked:

• The EXIT trap's own last command became the script's exit status, so a
  rm returning 0 turned a fatal abort into success. The trap now
  captures the status first and re-exits with it.
• bash 3.2 - which macOS ships - treats "${@}" and "${array[@]}" as
  unbound under set -u when empty, so passing an EMPTY array to
  expect_contains aborted the shell. Fixed in the library and in the
  tests that exercise it.

The suite now genuinely runs 10 of 10 rather than 4. A test framework
reporting green for tests it did not run is the worst bug it can have, so
this is the more important half of the change.

Verified: all 5 targets pass; skip counters asserted after test_runner
(a skipped test cannot assert about itself); the body-continues-past-skip
case fails loudly if the subshell ever goes away; --no-skip exercised
through the real flag parser, not just the variable.

Note on the version: bumped 0.5.1 -> 0.6.0. The subshell means a test body can no longer leak variables into later tests, which is a behaviour change even though it is the safer semantics.