fix(test): avoid restarting finished processes in batch subtest

[okurz]

Fix 1: t/02_parallel.t 'batch' subtest race (motivation for this PR)

CI for #83 failed on t/02_parallel.t line 74 (macOS, latest Perl):

#   Failed test at t/02_parallel.t line 74.
#          got: undef
#     expected: 'Hello world 3\n'

Root cause: after adding a 3rd process, $c->start() was called on the whole
Pool, which proxies to every member, re-forking the 2 already-finished
processes and racing with the new process's output. Fixed by using
$c->last->start() to start only the newly added process (matching the
pattern already used later in the same subtest).

Fix 2: t/13_shared.t semaphore-creation race

After fix 1, CI still failed in t/13_shared.t with pid counts off from the expected 20/21.

Root cause: Shared::Semaphore::_create() looks up an existing semaphore and,
if missing, creates one with IPC_CREAT|IPC_EXCL. When several forked
processes race to create the same semaphore concurrently, the losing
process's create call fails with EEXIST and confess()s, killing that
child instead of just attaching to the semaphore the winner created.

Fix: on a failed create, fall back to a lookup instead of failing (optimistically create first, fall back on failure to attach).

Fix 3: t/13_shared.t shared memory attachment resource leak

After fix 2, some runs under high resource contention / stress still failed in t/13_shared.t with shm_nattch > 0 assertions.

Root cause: IPC::SharedMem objects do not automatically detach from their shared memory segments upon garbage collection when they go out of scope. They only detach if detach() is explicitly called. This resulted in a cumulative resource leak of active memory attachments in the parent process from previous subtests (such as free_mem or manual locks), causing shm_nattch == 0 assertions to fail.

Fix: Added an explicit detach() call wrapped in eval inside Memory::DESTROY, ensuring all active attachments are cleaned up automatically as soon as a shared_memory helper instance goes out of scope.

Testing

• perl -Ilib t/02_parallel.t passes repeatedly.
• TEST_SHARED=1 perl -Ilib t/13_shared.t passes repeatedly (under extreme concurrent load + background stress testing).
• TEST_SHARED=1 TEST_SUBREAPER=1 prove -l t (full suite) passes repeatedly.
• perltidy reports no formatting changes needed for modified files.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.61%. Comparing base (5831e64) to head (463ec35).

Additional details and impacted files

@@           Coverage Diff           @@
##           master      #84   +/-   ##
=======================================
  Coverage   91.60%   91.61%           
=======================================
  Files          45       45           
  Lines        2848     2849    +1     
=======================================
+ Hits         2609     2610    +1     
  Misses        239      239           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[okurz]

refactored code to ensure full coverage

[Martchus]

This sentence isn't really comprehensible. What is "it" referring to? Maybe:

Suggested change

	# Try creating it. IPC_EXCL in flags fails with EEXIST if it already
	# Try acquiring a possibly existing semaphore using the IPC_EXCL flag. This will fail with EEXIST if it already

[Martchus]

Although it would still be nice to improve the wording as mentioned before.