Add Token statistics API

Support Hugging Face and ModelScope downloads for the Orin model and runtime, fixing the ModelScope-only source selector.

Improve Qwen 3.8 Flash Next performance on X3; single-stream decoding reaches 57.48 TPS.

Fix Pi Agent invocation errors.

X3 single-stream decoding improves by 7%, 8-way concurrent decoding improves by 61%, and 8-way concurrent TTFT is reduced by 93.5%.

Update Qwen 3.8 27B performance on X3