feat(ai): answer questions and summarize PDFs entirely on-device #77

Add a local document-assistant view grounded in the PDF currently selected by the user.

Scope

• Read an existing text layer first; offer OCR explicitly when one is missing.
• Produce answers and summaries with page references that open the relevant page.
• Keep extracted text, chunks, prompts, and responses on the device.
• Bound memory use for long documents and make cancellation immediate.
• State when the document does not contain enough evidence for an answer.
• Do not index folders or retain document content by default.

Acceptance criteria

• Works with the local runtime from v0.6 and with networking disabled.
• Page references are traceable to extracted source text.
• Scanned, mixed-language, empty, encrypted, and very large PDFs have defined behavior.
• Clearing the session removes its transient document data.