#2289 Anthropic API: cache_creation_input_tokens is never reported, though every newly read token is kept

说明

## Summary

`/v1/messages` never reports `cache_creation_input_tokens`. Every token a request reads is kept for the next request (the live session, and the conversation cache when it is on), which is what Anthropic calls a cache write. Clients that show cache writes, such as Claude Code, always see none.

## Expected

Anthropic's usage splits the prompt three ways, and the three add up to the whole prompt:

- `cache_read_input_tokens`: read from the cache;
- `cache_creation_input_tokens`: written to the cache by this request;
- `input_tokens`: neither.

## Actual

`message_delta`'s usage has `input_tokens` (everything read now) and `cache_read_input_tokens` (the reused prefix). `cache_creation_input_tokens` is absent, so a client fills in 0, while the tokens read now are in fact kept and reused by the next request.

## Impact

Display only: the totals and a client's context fill are already right. Claude Code's status line and cost estimate show every newly read token as uncached input, and no cache writes at all.

## Fix

#2290: the newly read tokens are `cache_creation_input_tokens`, all but the last prompt token, which stays `input_tokens` (never 0: Claude Code takes 0 as "not given" and counts the prompt twice, #1013).