fix(langchain): anthropic cache token count

[RogerHYang]

resolves #2381

---

Note

Fixes token accounting for Anthropic (including cache read/write) and adds parsing of LangChain UsageMetadata with detailed token breakdown; updates tests and bumps langchain_core minimum.

• Instrumentation (LangChain tracer):
  • Parse langchain_core.messages.ai.UsageMetadata to emit LLM_TOKEN_COUNT_* (incl. audio, reasoning, cache read/write) with Bedrock Anthropic heuristics.
  • Support raw Anthropic usage with cache_read_input_tokens/cache_creation_input_tokens; compute prompt/completion and cache detail attributes.
  • Refactor _token_counts to use new detectors/extractors while retaining OpenAI/VertexAI paths.
• Tests:
  • Update Anthropic token count expectation (prompt now includes cache; 33).
  • Add tests/test_token_counts.py with parametrized cases for UsageMetadata and raw Anthropic usage.
• Dependencies/Tooling:
  • Bump langchain_core min version to >=0.3.9 (and type-check pin).
  • Tox: add uv pip list -v in setup.

^{Written by Cursor Bugbot for commit 8f3b955. This will update automatically on new commits. Configure here.}