test: add LiteLLM Proxy runtime smoke test

[rlippmann]

Summary

Adds an opt-in Tier 2 runtime smoke test for the LiteLLM Proxy reference integration.

The test launches a real LiteLLM Proxy subprocess, loads the existing Context Compiler pre-call hook, sends requests through /v1/chat/completions, and verifies blocking and forwarding behavior against a local OpenAI-compatible stub upstream.

Why

Existing LiteLLM Proxy tests covered hook behavior directly, but did not prove that the real proxy runtime could:

• start with the integration hook configured
• load the callback
• execute the hook during request handling
• block confirmation-required requests before upstream forwarding
• forward allowed requests with the injected compiler contract
• shut down cleanly

This adds opt-in runtime evidence for the most important remaining LiteLLM Proxy gap tracked by issue #13.

Changes

• Added python/tests/test_litellm_proxy_runtime.py
• Added a proxy_runtime dependency group for litellm[proxy]
• Documented the opt-in runtime smoke command in the LiteLLM Proxy README
• Clarified callback import-path behavior for runtime configs

Scope

This PR does not:

• change hook behavior
• add a new example
• require OpenAI or Anthropic keys
• add the runtime smoke test to canonical validation
• close issue test: track remaining OpenWebUI runtime loading and registration validation gap #13 by itself

Notes

The runtime smoke test is opt-in:

RUN_LITELLM_PROXY_RUNTIME=1 uv run --group proxy_runtime pytest python/tests/test_litellm_proxy_runtime.py

The test uses a local OpenAI-compatible stub upstream rather than a live model provider.

Issue #13 impact

This partially addresses issue #13 by adding automated runtime coverage for LiteLLM Proxy startup, hook loading, request blocking/forwarding, and teardown.

Issue #13 should remain open until maintainers decide whether this satisfies the LiteLLM Proxy runtime requirement and until the remaining OpenWebUI or other runtime validation gaps are reviewed.