[Front-end] [Messages] Populate num_cache_creation_tokens

[yzong-rh]

Purpose

Follow up to #40912.

When --enable-prompt-tokens-details is toggled, compute the number of prompt tokens written to prefix cache and pass it through to OpenAI Chat Completions and eventually to Anthropic Messages API. Ensures num_cache_creation_tokens is populated.

Test Plan

.venv/bin/python -m pytest \
  tests/entrypoints/anthropic/test_anthropic_messages_conversion.py \
  tests/entrypoints/anthropic/test_messages.py::test_anthropic_streaming_cache_usage \
  tests/entrypoints/openai/chat_completion/test_serving_chat.py::test_mm_prompt_tokens_details

Send and examine real requests to Qwen3.6-35B-A3B served via Messages API, both streaming and non-streaming

Test Result

49 passed, 16 warnings in 35.46s

First request:
{
  "cache_creation_input_tokens": 4224,
  "cache_read_input_tokens": 0,
  "input_tokens": 527,
  "output_tokens": 197
}

Same request sent again:
{
  "cache_creation_input_tokens": 0,
  "cache_read_input_tokens": 4224,
  "input_tokens": 527,
  "output_tokens": 310
}

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[yzong-rh]

cc @tlrmchlsmth @zhangshuoming990105 @bbrowning

[tlrmchlsmth]

This is going to be too expensive to do for every request since it hashes every block

One potential alternative is to add a high-watermark number of blocks cached by the request to PrefillStats and then update it inside cache_blocks() in the KVCache coordinator implementations.

Otherwise if that doesn't work out, we can guard this by only calling find_longest_cache_hit when --enable-prompt-tokens-details is set but that will require plumbing

[yzong-rh]

Updating it inside cache_blocks() was quite messy. Found a reasonable way to estimate it given block.block_hash_num_tokens.

[yzong-rh]

In the current code, kv_transfer_params and ec_transfer_params will never be truthy without stopped being true.

[yzong-rh]

We take prefill_stats here to ensure we can traverse request's KVCacheBlocks before they are freed

[yzong-rh]

Two failures are v1/core/test_scheduler.py::test_abort_request_when_structured_output_fsm_cannot_advance and v1/core/test_async_scheduler.py::test_abort_request_when_structured_output_fsm_cannot_advance.

Caused by not mocking scheduler.kv_cache_manager.estimate_cached_tokens.return_value.