docs: write 'First Benchmark Run' step-by-step walkthrough

[JacobPEvans-personal]

Why

README.md shows a 3-step CLI block (lm-eval → dry-run → publish) but jumps over real-world setup questions:

• How do I bring up vllm-mlx + llama-swap? (link to nix-ai)
• What does max_gen_toks=4096 mean vs other gen_kwargs?
• How do I handle slow models / timeouts?
• What's the difference between dry-run and a real publish?
• How do I publish a partial run if my model crashed mid-suite?

Suggested approach

New docs/first-benchmark-run.md that walks through a complete first run end-to-end, with explanation:

1. Set up the inference stack (link to nix-ai).
2. Verify model availability (curl localhost:11434/v1/models).
3. Pick a task and run lm-eval with limit=10 (cheap smoke).
4. Inspect results_*.json (what the converter will see).
5. Dry-run publish, read the validation output.
6. Real publish, view the shard on HF dataset.
7. Open the HF Space viewer to see it.

Add a "Common pitfalls" section: wrong base_url (must end in /v1/chat/completions), unavailable model in catalog, missing HF_TOKEN.

Acceptance

• A new user with a working vllm-mlx setup can complete a real benchmark publish using only this doc.
• Cross-linked from README "Usage" section and from docs.jacobpevans.com/tools/mlx-benchmarks.

Filed as follow-up to the docs(presentation): polish work in #49.

[jacobpevans-claude]

🤖 Issue Solver attempted this issue and stopped.

Reason: cannot execute GitHub Contents API writes — gh api mutations require user approval in this interactive session; routine is designed to run in GitHub Actions where commands are auto-approved
Phase reached: implement

Human review needed. The agent will not retry this issue for 7 days.