[bowenzhu21]

Summary

• add a stock TensorRT 10 enqueueV3 executor
• use persistent CUDA input/output buffers and a non-blocking stream
• validate the initial input_ids/logits tensor contract
• add a TensorRT discovery module and compile/link smoke target
• compile in NVIDIA's documented TensorRT 24.08 development container

Validation

• CPU correctness, trace replay, artifact, and sanitizer suites pass locally
• PR CI will compile and link the actual TensorRT/CUDA path without requiring a GPU

Scope

This establishes the stock host-sampling baseline. GPU execution, numerical parity, KV bindings, and measured performance remain separate gated milestones.