Allow re-running an extraction on the same input

[chetanr25]

Allow re-running an extraction on the same input

One input gives one extraction, so testing the extractor meant re-uploading the same narrative every time.

Why

• A run that dies on a bad model name or a 503 from the provider burns the input.
• So does every prompt tweak. New upload, new input id, every time.
• The one extraction per input rule is right for a real deployment. It is only in the way during development.

What changed

• EXTRACTION_ALLOW_RERUN in app/core/config.py, default false.
• The route skips the existing extraction check when it is on.
• Documented in docker/.env.example, shipped as false.
• Two tests: a rerun returns 202 and leaves the earlier extraction alone, and the 409 still fires with the flag off.

Closes #674