Add GPT-6 Astra extract one-shot pipeline

[boyang-zhang1]

What

Adds the GPT-6 Astra one-shot structured-output extract pipeline and its leaderboard entry.

Why

Benchmark GPT-6 Astra on ExtractBench alongside the existing one-shot LLM/VLM extractors.

How

• Register openai_gpt_6_astra_reasoning_low_extract_oneshot_structured_output_file (provider openai_extract, reasoning_effort: low), matching the existing one-shot sibling config shape.
• Wire reasoning_effort through the OpenAI Responses provider — forwarded to the API as reasoning.effort only when set, so non-reasoning pipelines are unaffected.
• Add gpt-6-astra pricing ($10/M input, $50/M output) to the provider pricing table.
• Add the leaderboard row (Overall 91.91; Short 97.22 / Medium 90.56 / Long 31.70) and regenerate the README tables.

Changes

• src/extract_bench/inference/pipelines/extract.py — new pipeline spec
• src/extract_bench/inference/providers/extract/openai_responses.py — pricing entry + reasoning wiring
• docs/pipelines.md — pipeline row
• leaderboard.csv — new row (rank 3)
• README.md — regenerated from the leaderboard

Testing

• uv run extract-bench pipelines | grep gpt_6_astra resolves the new pipeline
• uv run ruff check passes on the changed files
• uv run python scripts/update_readme.py — diff confined to the generated table blocks

Notes

Requires OPENAI_API_KEY (no new env var). Grounding columns are 0.00 — one-shot structured output emits no citation boxes.