Replace source analysis with chunked resumable extraction

[justinjoy]

Summary

• replace monolithic web source analysis with durable extraction jobs and persisted source chunks
• process uploads chunk-by-chunk with progress, failed-chunk retry, resume on server restart, and duplicate chunk-claim protection
• link chunked candidate facts to source, run, and extraction job; clean jobs/chunks on source deletion
• preserve source text more strictly and drop translated Han facts from Korean sources
• make Ollama request timeout configurable while removing source analysis dependence on one long request

Tests

• .venv/bin/python -m pytest -q

Closes #70