Add Qwen3.8-Flash-Next-FP8 one-shot extract pipeline

[boyang-zhang1]

What

Adds qwen3_8_flash_next_fp8_vllm_extract_oneshot_structured_output_file, a self-hosted vLLM one-shot structured-output extract pipeline for Qwen3.8-Flash-Next-FP8, and its leaderboard result.

Why

Qwen3.8-Flash-Next is a strong open-weight model; this lands it on the ExtractBench leaderboard alongside the other self-hosted Qwen entries so the family is comparable end to end.

How

• Registers the pipeline via _pipeline_spec with provider_name="vllm_extract", mirroring the existing qwen3_8_27b_fp8_vllm_extract_oneshot_structured_output_file entry: json_object mode (structured_output: False), 32k output budget, 1h timeout, additional_properties_false.
• The endpoint is read from QWEN3_8_FLASH_NEXT_SERVER_URL (via endpoint_env_var); there is no default endpoint, consistent with the other open-weight pipelines.
• Regenerates the README leaderboard tables from leaderboard.csv with scripts/update_readme.py.

Changes

• src/extract_bench/inference/pipelines/extract.py — new pipeline registration.
• .env.example — QWEN3_8_FLASH_NEXT_SERVER_URL entry.
• docs/pipelines.md — Extract Pipelines table row.
• leaderboard.csv — one new row (Overall 89.88; Short 94.82 / Medium 87.81 / Long 37.74).
• README.md — regenerated leaderboard/grounding tables.

Testing

• uv run extract-bench pipelines lists the new name.
• uv run ruff check passes on the changed file.
• uv run python scripts/update_readme.py reproduces the README tables from the CSV with no manual edits.

Notes

• Self-hosted open-weight model: set QWEN3_8_FLASH_NEXT_SERVER_URL to your vLLM deployment before running; cost columns are blank and grounding columns are 0.00 (json_object mode emits no bounding boxes), matching the sibling Qwen rows.