Fix Pandoc subdocument XML namespace handling - #658

Requesting changes for two blocking issues:

1. Subdoc._get_xml() now serializes each direct body child independently. lxml writes every inherited namespace declaration onto every child, including namespaces that are not used by that child or are already available from the destination document. In a CPython 3.12.3 / lxml 6.1.1 benchmark with a 10,000-paragraph dynamic subdocument, the intermediate fragment grew from 340,000 bytes at 1f143fb to 11,800,000 bytes at 62e0075; median render() time across three fresh runs grew from 0.083 s to 2.400 s. The same path affects every subdocument, not only Pandoc input. Please preserve only the missing or conflicting declarations needed by the inserted content and add a scale regression.

2. tests/pandoc_subdoc.py is not exercised by the current test job. The job invokes python tests/runtests.py from the repository root, but the runner globs [A-Za-z]*.py relative to the current directory, so it runs setup.py rather than the scripts under tests. It also ignores each subprocess.call() return code. Against 1f143fb, the new regression script exits 1 with the reported namespace error, while runtests.py still exits 0. Please discover tests relative to file and propagate child-process failures, or invoke this regression directly from the test job.

The namespace fix itself handles the reported document correctly: the Pandoc 3.9 reproduction fails at the base with the unbound a prefix and succeeds at this head with the image relationship, a/pic elements, nested section properties, and one final body sectPr intact. Custom default namespaces, prefixed attributes, and conflicting prefix mappings also remained resolved.

I rechecked exact head 24e1c2f after the serialization and test-gate changes.

• The #620 Pandoc fixture now preserves the DrawingML/PictureML nodes and image relationship, reopens with python-docx, and leaves exactly one final body sectPr.
• The default-namespace, namespaced-attribute, and conflicting-prefix cases remain resolvable.
• The 10,000-paragraph regression retains all 10,000 paragraphs in a 340,000-byte fragment, avoiding the prior per-child namespace amplification.
• The root-invoked runner discovers and executes pandoc_subdoc.py and propagates failures. I also ran the related subdocument, merge, and header/footer scripts and an exact-venv CLI render separately.
• Flake8, sdist/wheel builds, and git diff --check passed.

I did not repeat the interactive Word GUI check, but I found no remaining blocker in the XML/package behavior or in the two previously requested changes.