Add Gemini 3.8 Flash extract one-shot pipeline

[boyang-zhang1]

What

Add Gemini 3.8 Flash as a one-shot structured-output extract pipeline.

Why

Gemini 3.8 Flash is generally available (2026-09-02) at the same introductory price as 3.7 Flash, so the leaderboard should track the latest Flash tier.

How

• Register gemini_3_8_flash_extract_oneshot_structured_output_file via the gemini_extract provider (one-shot structured output over the uploaded file, thinking_level: low).
• Add gemini-3.8-flash to the Gemini extract pricing table ($0.75 input / $3.75 output per 1M tokens).

Changes

• src/extract_bench/inference/pipelines/extract.py — new pipeline spec
• src/extract_bench/inference/providers/extract/gemini_direct.py — pricing entry
• docs/pipelines.md — pipeline table row
• leaderboard.csv / README.md — leaderboard row and regenerated tables

Testing

• uv run extract-bench pipelines | grep gemini_3_8_flash_extract_oneshot_structured_output_file — registers.
• uv run ruff check on the changed files — passes.
• Benchmarked on the full 370-document set: Overall 80.71 (Short 88.14 / Medium 72.23 / Long 28.70). Word and page grounding are 0.00 — a values-only one-shot returns no citation boxes, same as the existing 3.5/3.6/3.7 one-shots.

Notes

Uses the existing Gemini API key env vars (GOOGLE_GEMINI_API_KEY / GEMINI_API_KEY / GOOGLE_API_KEY); no new configuration.