### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

When a Responses API request is routed through the `litellm_completion` transform (used for Bedrock/Anthropic among others, since they lack a native Responses API integration), any `cache_control` set on an input item gets dropped while the message is being converted to Chat Completion format. This silently breaks prompt caching for multi-turn or agentic sessions that go through this bridge - no error, caching just quietly stops working.

Expected: `cache_control` should be preserved through the transform so prompt caching still works.

Affects three message-building paths: plain user/assistant messages, tool call outputs, and function calls. For user/assistant messages specifically, `cache_control` also needs to land on the last content block, since Anthropic/Bedrock read the cache breakpoint at the content-block level rather than the message level.

### Steps to Reproduce

1. Send a multi-turn `/v1/responses` request to an Anthropic or Bedrock model with `cache_control` set on an input item 2. Inspect the outbound request sent to the provider…