### Check for existing issues

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

### What happened?

For Anthropic Claude models called over the Bedrock Invoke streaming path (bedrock/invoke/…), the final chunk carries token usage in amazon-bedrock-invocationMetrics. LiteLLM maps inputTokenCount and outputTokenCount from that block into usage, but silently drops cacheReadInputTokenCount and cacheWriteInputTokenCount.

As a result, prompt-cache usage never reaches usage.cache_read_input_tokens / usage.cache_creation_input_tokens, and the cost calculator bills cache-heavy traffic (e.g. a client with a large cached system prompt) as if every request were fresh input — a 4–7× over-report of input cost and a total under-count of cached tokens.

Root cause is in two places on the Invoke path:

1. Decoder — AmazonAnthropicClaudeMessagesStreamDecoder._chunk_parser in litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py. The amazon-bedrock-invocationMetrics mapping reads only inputTokenCount and outputTokenCount; the two cache fields are never read. This is distinct from _promote_message_stop_usage (LIT-2…