Refactor scraping module into layered source/pipeline/crawler packages

[MaxDall]

The scraping/ module had grown into a few large files (crawler.py ~870 lines, scraper.py, html.py) that mixed fetching, parsing, and orchestration. This PR splits them along the three-layer flow from the architecture docs — Source → Pipeline → Crawler — so each layer is its own module and independently testable. No public API change: everything re-exported from fundus/__init__.py stays the same.

Source layer — scraping/pipeline/source/

Raw HTML producers behind the HTMLSource protocol:

• web.py — WebSource for live-web fetching, with an internal _Pacer that enforces per-host crawl delays and WebSourceInfo carrying per-source metadata.
• ccnews.py — CCNewsSource reading the CC-NEWS WARC archive (WarcSourceInfo, with WarcFileLoadError for unreadable segments).

Pipeline layer — scraping/pipeline/__init__.py

Pipeline pairs a source with the publisher's parser, turning each HTML into an Article or dropping it via the extraction/language filters. Owns the HTMLSource protocol and PipelineError.

Crawler layer — scraping/crawler/

Public entry points, split by backend:

• base.py — CrawlerBase (the crawl() contract) and the internal _CrawlState (progress/limit bookkeeping).
• web.py — Crawler, threaded across publishers, plus the publisher_context_wrapper that scopes per-publisher cancellation state.
• ccnews.py — CCNewsCrawler, multiprocessing over the archive.
• queueing.py — shared producer/consumer plumbing (enqueue_results, iter_pool_results) and RemoteException for surfacing worker-process errors.

Cross-cutting

• Utils: new concurrency.py (execution-context detection, dill_wrapper for pickling closures across processes, a proxied tqdm for multiprocessing progress) and timing.py (random_sleep); timeout.py rewritten to compose a ResettableTimer instead of subclassing threading.Thread; serialization.py and events.py refactored with fuller docstrings.
• Primitives html/article/url/filter refactored, with supporting cleanup in parser/ and publishers/base_objects.py.

Fixed / improved

• 5xx retry in the session layer (session.py): the stack previously had no retry anywhere, so any transient 5xx aborted the source. 5xx responses are now retried in place with interruptable full-jitter exponential backoff (configurable via max_retries / retry_backoff_base / retry_backoff_cap, honoring Retry-After); an exhausted retry surfaces as a normal HTTPError.
• timeout hardening: dropped the fragile threading.Thread subclass that reached into private _target/_args (with a type: ignore) and the while True and … loop; Timeout(None) now cleanly disables the timer instead of relying on the separate disable= flag.

Tests

Reorganized to mirror the source tree (tests/scraping/{crawler,pipeline,…}, tests/utils/, tests/parser/, tests/publishers/), with shared fixtures/builders under tests/fixtures/ and a tests/README.md documenting conventions. New unit coverage for the timeout, serialization, queueing, and source/pipeline layers. An integration pytest marker separates the mocked-I/O end-to-end tests.

TODO (this PR)

• Refactor logging — loggers currently propagate upstream, producing duplicate (double-printed) messages.
• Fix Publisher.__eq__ — it compares self.parser by identity and ParserProxy defines no __eq__, so two value-equal publishers never compare equal. The bug is currently pinned by a strict xfail in tests/publishers/test_base_objects.py; remove the xfail once fixed.

Outlook (follow-up PRs)

• Replace the global __EVENTS__ registry with explicit CancellationToken objects (planned redesign documented at the top of events.py). The current registry conflates cancellation, shutdown propagation, and post-mortem queryability into one string-keyed, thread-id-resolved mechanism, has no clean seam for multiprocessing (threading.Event doesn't cross process boundaries), and forces __EVENTS__.context("test") boilerplate into every source test.
• Refactor the parser / publisher module — publishers/base_objects.py was only lightly touched here; the parser-proxy and publisher layer still deserve the same restructuring/test treatment.
• Adopt a TypedDict in article.py once PEP 728 lands (tracked by an inline TODO).