Fix resource leaks, Perl interpreter lifecycle, and silent failures

[quinnjr]

Summary

Applies findings from a combined /simplify + /optimize + /audit pass, followed by a full multi-dimension code review of the resulting diff, then adds test coverage for the gaps that review surfaced.

• Perl (src/languages/Perl.cxx, Perl.h): the interpreter is now constructed once and reused across plugin calls instead of a full alloc/parse/destruct per invocation (large perf win on pipelines with multiple Perl steps). Load failures now throw and are caught by main.cxx's existing fail-fast path instead of crashing the process or silently reusing the previous plugin's input/run/output subs.
• PerlIO.pm: fixed a two-argument open() (shell-injection vector via a crafted filename), a scalar-context arg-parsing bug, a wrong-sigil return that made readParameters() always return nothing, and a missing trailing 1; (Perl modules must return true). Also flagged separately (not fixed here, needs a maintainer call): the module's name collides with Perl's own built-in PerlIO I/O-layer pragma, so use PerlIO; in a plugin silently resolves to the wrong module — will file as a separate issue.
• src/languages/Compiled.cxx: fixed a leaked Plugin*/dlopen handle on every plugin invocation; mutex-guarded the new dlopen handle cache against concurrent LitterLaunch pipelines.
• src/languages/Py.cxx: mutex-guarded the new sys.path dedup cache for the same concurrency reason; fixed a new[]/delete mismatch.
• src/main.cxx: a plugin exception or "no suitable language for plugin" now aborts the pipeline (previously both were silent no-ops that let the run continue and exit 0). exit(1) no longer fires directly from a LitterLaunch worker thread — sibling threads get to finish first via a shared atomic flag checked after joins.
• src/PluginManager.h: create() no longer null-derefs on an unregistered plugin name; supportedLanguages() is now properly gated by HAVE_PYTHON/HAVE_PERL/HAVE_R so --without-* builds actually link.
• SConstruct: --without-{python,perl,r} now also gates compilation of the corresponding language source and linked libraries (previously only affected some targets, causing link failures); removed a hardcoded, machine-specific JVM library path left over from local debugging.
• src/languages/R.cxx: fixed a dangling-pointer trap between unload()/load().
• Assorted: PluGen generator bugs (undeclared Java field reference, OOB vector access, substring-vs-exact plugin-name matching), testPluMA.py's check() returning True after only checking the first line of a diff, Dockerfile layer ordering for build caching, dead-code removal, deduplicated the checkPool.py/getPool.py/getPlugins.py plugin-pool scraping logic into a shared pool_utils.py.

Test coverage added

• plugins/PerlChainStepA + plugins/PerlChainStepB: two minimal Perl plugins chained in one config.txt, the first example anywhere in the repo that runs more than one Perl plugin per process — exercises the interpreter-reuse code above. Verified locally: both plugins execute in one process, each produces its own correct output (not stale/leaked from the other), matching .expected files exactly.
• tests/test_negative_paths.sh (wired into CI): verifies ./pluma exits non-zero when a config references an unregistered plugin name, instead of the previous silent success. Verified locally against a live build.

Test plan

• Full project build (scons, and scons --without-r) with no new compiler errors in any touched file
• All touched Python files pass py_compile; PerlIO.pm and both new plugin scripts pass perl -c
• Manually built ./pluma and ran both new test fixtures against it directly — both pass
• CI (testPluMA.py + the new negative-path test) — will run automatically on this PR