docs(anthropic): clarify prompt caching middleware docstring (#38206)

Updates the `AnthropicPromptCachingMiddleware` class docstring so it no
longer implies the middleware itself tags the final message tail. It
tags the system message and tool definitions and passes `cache_control`
via `model_settings`; the chat model/provider applies the message-tail
and provider-specific behavior. Docstring-only, no runtime change.

Made by [Open
SWE](https://openswe.vercel.app/agents/27913c08-b40c-015e-afee-cf66788b7f08)

---------

Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>

libs/partners/anthropic/langchain_anthropic/middleware/prompt_caching.py

@@ -39,14 +39,17 @@ class AnthropicPromptCachingMiddleware(AgentMiddleware):
 
     Requires both `langchain` and `langchain-anthropic` packages to be installed.
 
-    Applies cache control breakpoints to:
+    The middleware tags stable agent content and passes `cache_control` through
+    `model_settings`:
 
     - **System message**: Tags the last content block of the system message
         with `cache_control` so static system prompt content is cached.
-    - **Tools**: Tags all tool definitions with `cache_control` so tool
-        schemas are cached across turns.
-    - **Last cacheable block**: Tags last cacheable block of message sequence using
-        Anthropic's automatic caching feature.
+    - **Tools**: Tags the last tool definition with `cache_control`. Because
+        tool definitions are sent as one contiguous block, a single trailing
+        breakpoint caches the entire tool set across turns.
+    - **`model_settings`**: Passes `cache_control` to the chat model. The chat
+        model/provider then applies the correct message-tail and
+        provider-specific behavior at request time.
 
     Learn more about Anthropic prompt caching
     [here](https://platform.claude.com/docs/en/build-with-claude/prompt-caching).