feat(node): readability main-content extraction

[gaborbernat]

Adds readability-style main-content extraction, the role resiliparse's main-content extractor fills.

• Node.main_content() -> Element | None returns the dominant content element (article body, with nav/sidebar/ads/comment boilerplate scored out), or None when nothing reads as content.
• Node.main_text() -> str renders that element with to_text(), or "".

Heuristic

A content-density score over the arena tree, entirely in C (treebuilder_readability.h):

• Paragraph-like elements (<p>, <td>, <pre>, >=25 chars) contribute a base point, one per comma (a clause proxy), and up to three for length (one per 100 chars); the contribution rolls up to the parent in full and the grandparent at half.
• Containers get a tag weight (<div> +5, <blockquote>/<td>/<pre> +3, lists/<form> -3, headings/<th> -5) and a class/id weight (+/-25 for content vs boilerplate hints).
• Boilerplate subtrees (<script>, <nav>, <aside>, foreign namespaces, and comment/modal/sidebar-class elements not rescued by a main/content hint) are pruned before counting.
• Each survivor is discounted by its link density; the highest remaining score wins.

The scoring walk is pure C and free-threading safe: it touches no Python object until a winner is chosen, then the binding wraps (or renders) that one node under the per-tree critical section. A concurrent stress test covers it.

Scope

Language detection and WARC/archive handling that resiliparse bundles are out of scope; the migration guide points to a dedicated tool for those.

Docs & tests

• Tutorial, how-to, reference (autodoc) and explanation coverage, plus the resiliparse migration row.
• New tests/test_tree_main_content.py drives every scoring branch (100% C line+branch under both llvm-cov and gcc-16); concurrent stress in test_tree_freethread.py.

closes #249