docs: explain benchmark-backed experiment workflows

[joshblack]

Closes #329.

This PR completes the benchmark-backed experiment workflow guidance for the stack. It adds configuration and CLI examples and explains the reference treatments, setup isolation, trial matrix, saved hierarchy, and capability-first results.

Changelog

New

• Add complete benchmark-backed experiment examples.
• Add guidance for comparing capability results and drilling into scenarios.

Changed

• Update benchmark, experiment, plan, treatment, and result references for the new workflow.
• Update trial-count guidance for the two automatic reference treatments.
• Clarify that grouping snapshots do not freeze setup code or scenario contents.

Removed

None.

Validation

• pnpm exec agent-eval experiment plan create benchmark-comparison --experiments .agents/tmp/doc-example/experiments --benchmarks .agents/tmp/doc-example/benchmarks --scenarios .agents/tmp/doc-example/scenarios --output-path .agents/tmp/doc-example/plan.json: passed; produced the expected three reference/custom treatment trials.
• pnpm run format:diff: passed.
• pnpm exec turbo run lint lint:npm: passed.
• pnpm exec turbo run test -- --run: passed.
• pnpm exec turbo run type-check --filter='!@primer/agent-eval-website': passed.

The website build was not repeated for this documentation-only entry. No live agent evaluations were run.