fix(probes,generators): raise GarakException on unpicklable parallel worker payload

[glatinone]

Summary of Changes

Probe._execute_all (garak/probes/base.py) and Generator.generate (garak/generators/base.py) each catch pickle.PicklingError around their multiprocessing.Pool.imap_unordered call, alongside the existing OSError/errno-24 handling, and re-raise it as a GarakException with actionable guidance to reduce parallel_attempts/parallel_requests to 1 — mirroring the existing "Parallelisation limit hit" pattern already in both functions.

Root Cause & Performance Impact

Both parallel_attempts (probes) and parallel_requests (generators) submit work to a multiprocessing.Pool via imap_unordered. Submitting a task pickles the bound method together with self (the probe/generator instance). If that instance holds an attribute that cannot be pickled — e.g. a dynamically loaded plugin object whose defining module can't be re-imported by name in the worker process — Pool._handle_tasks raises a raw _pickle.PicklingError from deep inside multiprocessing internals, exactly as shown in the two tracebacks in #361 (import of module 'core.py' failed, import of module 'summarize_document.py' failed).

This isn't a performance issue so much as an abrupt, unhelpful failure mode: the run aborts with an internal multiprocessing traceback instead of a clear, actionable error — the maintainer's own framing on the issue was "what should we do? at least explode /politely/, for now." This PR scopes strictly to that: converting the raw internal traceback into a clear GarakException. Auto de-parallelization (mentioned as a follow-up in the issue) is intentionally left out of scope.

Test Coverage & Verification

• tests/generators/test_generators.py::test_parallel_requests_unpicklable_attribute_raises_garak_exception — reproduces the generate() path (parallel_requests), using a module-level stub function whose __module__ is corrupted to force a genuine pickle.PicklingError (import of module '...' failed) matching the issue's exact failure shape, rather than relying on a local-closure AttributeError.
• tests/test_internal_structures.py::test_execute_all_unpicklable_probe_raises_garak_exception — same technique for the Probe._execute_all path (parallel_attempts).
• Both tests were verified to fail against the pre-fix code with the exact _pickle.PicklingError: Can't pickle ...: import of module '...' failed traceback from fail more helpfully when parallelism not supported #361, and to pass after the fix (asserting GarakException is raised instead).
• Full suite: pytest — 1489 passed (only the 2 directly touched test files plus tests/test_config.py -k parallel were run against both pre- and post-fix code to isolate signal). One pre-existing failure (test_probe_metadata[probes.audio.AudioAchillesHeel], missing optional soundfile/librosa extras) and 3 pre-existing errors (openai_compat_mocks fixture not resolving when test_openai.py is invoked alongside certain other test paths) were confirmed present identically on main before this change — unrelated to this diff.
• black --check and pylint show no new issues introduced by this diff (both tools flag two pre-existing, unrelated formatting/lint items on main in the touched files, outside the lines this PR changes).

Closes #361