Add Token statistics API

Deploy Qwen3-Embedding 0.6B, 4B and 8B on AGX Orin (JetPack 6.2 / 7.2) and AGX Thor (JetPack 7.2) with an OpenAI-compatible embeddings API. Supports query instructions, flexible dimensions, base64 encoding and 32K context. Inference uses CUDA BF16 / SDPA, with bounded CUDA Graph optimization for short single-input 0.6B requests. Download runtime archives from immutable Hugging Face / ModelScope snapshots with size and SHA256 verification. Model weights are downloaded separately and loaded offline. Includes nine native GPU benchmark groups (1,800 requests), bilingual retrieval checks and deployment lifecycle validation across three platforms. Performance scope and stability limitations are documented. Deployment from the AI Pod interface requires a main application version that includes the Qwen3-Embedding model card. That card is currently absent from the test box's main application; the interface dependency is tracked separately.