feat(examples/appworld): add AppWorld RLM benchmark

[glesperance]

Rationale

This adds the AppWorld benchmark path for predict-rlm so we can evaluate RLM agents on realistic app/API tasks, run grouped GEPA optimization, and publish sanitized benchmark results without committing raw traces or private run artifacts.

Summary

• Add an examples/appworld package with benchmark loading, AppWorld worker tools, PredictRLM service wiring, GEPA project config, smoke tests, and setup helpers.
• Add grouped AppWorld optimization support in rlm-gepa, async evaluation adapter support, and related reporting/statistics utilities.
• Harden sandbox/interpreter behavior around async stderr EOF hangs, defaulted single-output SUBMIT, and tool kwarg handling.
• Rename repo agent guidance from CLAUDE.md to AGENTS.md and add long-running local run guidance.
• Refresh docs, generated skill guidance, example outputs, and AppWorld blog/results documentation.
• Keep raw AppWorld run artifacts ignored; document the externally hosted protected runs.bundle instead.

Test Plan

• AppWorld smoke coverage added under examples/appworld/tests/test_appworld_smoke.py.
• Unit coverage added/updated for interpreter behavior, PredictRLM submit handling, LM config, RLM skill docs, and GEPA runtime/reporting.
• AppWorld benchmark results recorded in results.txt with zero errors/timeouts for the top GPT-5.5 run: 154/168, TGC 0.917, SGC 0.839.
• Raw run artifacts are excluded from git; only examples/appworld/runs/README.md is tracked.