fix: harden long PDF page extraction

[gwokhou]

Summary

This PR hardens long PDF page extraction in index_long_document().

It normalizes page content returned by PageIndex Cloud or the local PDF fallback into OpenKB's expected source JSON shape, including support for common page fields such as page,
page_number, page_num, content, markdown, and text.

Why

Long PDF indexing currently assumes that page extraction returns data in the exact shape OpenKB later writes to wiki/sources/*.json. In practice, cloud/local extractors may return
strings, alternate page-number fields, alternate content fields, invalid image metadata, or empty/unusable page data.

That can lead to brittle downstream behavior during wiki compilation, especially for complex or lengthy PDFs.

Related to #77. This does not replace the default PDF parser, but it improves the resilience of the existing PageIndex/local PDF extraction path.

Changes

• Add _normalize_page_content() for PageIndex/local PDF page outputs.
• Normalize cloud get_page_content() responses before writing source JSON.
• Normalize local PDF fallback output as well.
• Fall back to local extraction when cloud page content is empty or invalid.
• Raise a clear RuntimeError when both cloud and local extraction produce no usable page content.
• Add tests for normalized page shapes, invalid cloud fallback, and empty extraction failure.

Testing

• Added focused unit coverage in tests/test_indexer.py.