Refresh BenchFlow docs with a local-first quickstart

[ElegantLin]

Summary

• replace the cloud-first onboarding path with a local Docker + subscription-auth flow
• add a tested Start in 5 minutes guide using the real SkillsBench 3d-scan-calc task
• add focused authentication, evaluation-running, and sandbox-selection guides
• refresh the README, concepts, examples, agent, CLI, and Python API documentation
• document local rubric review and fix the removed standalone skill-eval example

Why

The existing docs assumed too much prior BenchFlow knowledge and made Daytona look necessary for ordinary local use. The public onboarding path should start with one understandable local evaluation, explain what pass/fail means, and defer cloud sandboxes until remote scale or isolation is actually needed.

Validation

• UV_CACHE_DIR=/private/tmp/benchflow-uv-cache uv run python -m pytest tests/test_cli_docs_drift.py — 8 passed
• real local Docker run of 3d-scan-calc with Codex/gpt-5.5 — 1/1, reward 1.0, 9 tool calls, complete timing/token metadata
• relative-link audit across all 22 public BenchFlow docs — passed
• git diff --check — passed

Companion website PR: benchflow-ai/www.benchflow.ai#55.