Expand provider adapters and packaging for the model runner

[cursor]

Summary

runner.py ships FakeModelRunner and stdlib OpenAICompatibleRunner. Teams need clearer extras packaging, richer limits, and at least one additional provider path without forcing SDKs on offline users.

Acceptance criteria

• Optional dependency extras (e.g. tool-semantics[openai]) documented; core install stays SDK-free
• Cost / token / max-call limits are enforced and recorded in RunnerMetadata.run_config
• At least one additional adapter or a documented generic HTTP profile matrix (OpenAI, Azure OpenAI, local OpenAI-compatible)
• Retry/backoff behavior documented and tested with fakes
• Unit tests use fakes only (no live network in CI)
• docs/probes.md updated with runner configuration examples

Non-goals

• Requiring Anthropic/OpenAI SDKs for default install