fix: take streaming token counts from the engine

[Priyanshu-u07]

Nothing asked upstream for usage on a stream, so every streaming request fell through to estimating tokens from the message text, which misses the chat template. Measured at 10 prompt tokens against the engine's 35. check_quota and increment_redis_only count against these numbers, and the Insights token totals are built from them.

StreamProcessor already preferred provider usage when it saw any — it never saw any. The request now sets stream_options.include_usage unless the client set stream_options itself, and the usage-only chunk that produces is recorded and
dropped rather than forwarded, since the client did not ask for it.

External providers are unaffected. The injection happens before adapter.transform_request, and AnthropicAdapter and CohereAdapter both rebuild the payload from scratch, so the field is discarded before it reaches those APIs. The Anthropic surface sets stream_options itself, so this leaves that path alone.

Verified on a live deployment — A10G, vLLM 0.22.1 serving Qwen2.5-14B-Instruct-AWQ. A streaming request for "Count from one to twenty." logged prompt_tokens 35, matching what the same prompt reports non-streaming. The response ends at finish_reason "stop" then [DONE], with no usage frame.

closes #381