Add Token statistics API

Initial ReaderLM v2 release for local HTML-to-Markdown and structured JSON on AGX Orin / AGX Thor. Thor automatically uses calibrated NVFP4 with CUTLASS; Orin uses Q4_K_M. Total input plus output context is fixed at 32K. Select a device and ModelScope or Hugging Face in AI Pod. Includes verified downloads, progress, cancellation, retry, stop and restart. Model and device-runtime archives are not uploaded to a CDN. Derived from Jina AI ReaderLM-v2 under CC-BY-NC-4.0, requiring attribution and non-commercial use. Requires an LZC AI Pod main application supporting the model deployment protocol.