[feat] url based rollout

[mathewjhan]

What

Why

How to Test

Checklist

• PR title follows [module] type: description format
• Appropriate labels added (e.g. enhancement, bug, breaking)
• ruff check . and ruff format --check . pass
• pyright osmosis_ai/ passes
• pytest passes (new tests added if applicable)
• Public API changes are documented
• No secrets or credentials included

---

Summary by cubic

Switch rollouts to URL-based routing and a single-sample contract to simplify execution, grading, and integrations. Adds minimal multiply examples for openai_agents and strands.

• New Features

  • URL carries identity via chat_completions_url and callback paths; rollout_id in the body is optional (server generates one if missing).
  • Single-sample API across backends and types; Harbor/local write sample.json and reward.json; grader callbacks send a single sample.
  • OsmosisLitellmModel no longer stamps per-call headers; models rely on URL routing.
  • New multiply examples for openai_agents and strands, each with a local server, tools, workflow, and grader.

• Migration

  • Replace dict-of-samples with a single sample: ExecutionResult.sample, GraderInitRequest.sample, GraderCompleteRequest.sample; RolloutSample no longer has id; remove MultiTurnMode.
  • Update context/graders: use RolloutContext.set_sample_source(...), get_sample(), and GraderContext.set_reward(...).
  • Harbor/local runners: use --sample and sample.json; grader writes reward.json as {"reward": <float>}.
  • Integrations: pass OsmosisMemorySession() to OpenAI Agents runners; use OsmosisStrandsAgent with OsmosisRolloutModel in Strands; models rely on URL routing.

^{Written for commit d55ebc8. Summary will update on new commits.}