### Check for existing issues

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

### What happened?

The OTEL callback's `_set_usage_outputs` function never sets reasoning token counts (`llm.token_count.completion_details.reasoning`) on spans for Chat Completions API responses, even when the Usage object contains valid `completion_tokens_details.reasoning_tokens`.

**Root cause:** Line 239 in `litellm/integrations/arize/_utils.py` looks up `output_tokens_details`:

```py reasoning_tokens = usage.get("output_tokens_details", {}).get("reasoning_tokens") ```

But Chat Completions API uses `completion_tokens_details`, not `output_tokens_details`. The `output_tokens_details` key only exists on `ResponseAPIUsage` (Responses API). Since `Usage.get()` delegates to `getattr()`, it returns the default `{}`, and `{}.get("reasoning_tokens")` is always `None`.

Additionally, `prompt_tokens_details` (cached_tokens, cache_creation_tokens, audio_tokens) are never extracted at all, even though the corresponding `SpanAttributes` constants exist.

This affects all OTEL-based callbacks that use `_set_usage_outputs`: Langfuse (OTEL mode), Arize, Phoenix, Weave.

#…