perf(provider): optimize anthropic prompt cache hit rate

[bitifirefly]

Summary

• backport upstream LiteLLM prompt-cache optimization for Anthropic-compatible routes
• improve _apply_cache_control to place two cache breakpoints:
  • system message
  • second-to-last message (conversation history prefix)
• keep tool cache marker behavior (last tool definition)
• update tests to verify both breakpoints and non-mutation of inputs

Validation

• ruff check opencane/providers/litellm_provider.py tests/test_litellm_prompt_caching.py
• pytest -q tests/test_litellm_prompt_caching.py tests/test_litellm_message_sanitize.py
• pytest -q (full): 396 passed