Anthropic document_index citations don't account for documents embedded in custom tool results

Context

Follow-up from #375-adjacent work in fix/anthropic-web-fetch-citations (anthropic_document_sources() in chatlas/_provider_anthropic.py), which resolves Anthropic document_index citations to source URLs across the whole request.

Gap

anthropic_document_sources() walks each turn's contents and assigns a document_index slot for ContentToolResponseFetch, ContentPDF, ContentDocument, and non-image ContentUploaded. It does not look inside ContentToolResult.get_model_value().

A custom tool can return ContentToolResult(value=..., model_format="as_is") whose value is a raw Anthropic block list containing a DocumentBlockParam or SearchResultBlockParam. That block is sent to Anthropic as part of the tool result content and occupies a document_index slot server-side, but anthropic_document_sources() never accounts for it — so every document_index computed for citations after that tool result is off, and citations pointing at the tool-embedded document can never resolve to a source.

Reproduction sketch

1. Register a custom tool that returns ContentToolResult(value=[{"type": "document", ...}], model_format="as_is").
2. Follow it with a web_fetch or another document-producing turn.
3. Citations for content after the tool result resolve to the wrong WebSource (or None) because the tool-embedded document's slot wasn't counted.

Notes

This is a narrow/advanced case — it requires a custom tool to hand-construct raw Anthropic document/search-result blocks (not part of chatlas's documented tool-result patterns) — so it was shipped as a known limitation rather than blocking the citation-resolution fix. Fixing it would mean extending anthropic_document_sources() to also scan ContentToolResult's raw content for document-shaped blocks.

Start in chatlas/_provider_anthropic.py at anthropic_document_sources() and inspect how ContentToolResult.get_model_value() is represented in the request contents. Reproduce the custom as_is tool result followed by a document-producing turn, then verify that embedded document and search-result blocks consume document_index slots and later citations resolve to the correct WebSource.