Add GLM-5.3 Flash extraction pipeline

[boyang-zhang1]

What

Adds the Z.ai GLM-5.3 Flash one-shot structured extraction pipeline and its public leaderboard results.

Why

This makes the hosted model directly runnable in ExtractBench and publishes its value quality, cost, and latency characteristics.

How

• Sends documents inline through Z.ai's OpenAI-compatible chat API.
• Embeds the requested JSON schema in the prompt and requests JSON object output.
• Accounts for standard and cached input tokens using public list pricing.
• Loads credentials from GLM_ZAI_API_KEY.

Changes

• Added the GLM Z.ai extraction provider and pipeline registration.
• Documented setup and the required environment variable.
• Added the benchmark row and regenerated the README leaderboard blocks.

Testing

• Full test suite: 788 passed.
• Ruff and mypy passed on the changed Python files.
• Pipeline discovery confirmed the new pipeline is registered.
• Mocked inference verified file input, JSON output, cached-token cost accounting, and normalization.
• README regeneration and git diff checks passed.

Notes

• Requires a Z.ai API key in GLM_ZAI_API_KEY.
• Unified value F1 is 80.75 overall: 96.30 Short, 51.56 Medium, and 27.83 Long.
• Average cost is 0.34 cents per page.
• The pipeline does not emit evidence citations, so grounding scores are zero.
• No breaking changes; the pipeline is opt-in.