feat(runtime): add native TensorRT deployment for SenseVoice - #3463

Summary

• add a native TensorRT builder for the SenseVoice encoder with ONNX contract validation, dynamic optimization profiles, FP16/FP32 selection, and atomic plan output
• reject dynamically quantized ONNX graphs before parsing, with the exact quantize=False recovery command
• add a Triton tensorrt_plan config template while preserving the existing ONNX Runtime default
• document export, build, profile sizing, plan portability, and verified H100 results

Verification

• red-first TensorRT contract tests: 10/10 passed
• merged-main focused suite (builder contracts plus related docs/runtime checks): 31/31 passed
• Black 24.4.0 and git diff --check: passed
• official FP32 ONNX: checker passed; TensorRT 10.0.1 parser returned 0 errors
• H100 FP16 build: 527,504,916-byte plan in 113.881 s, profile batch 1-16 / frames 1-4096
• runtime boundaries: (1, 1), (8, 512), and (1, 4096) produced finite outputs
• random 30/64-frame parity: 100% CTC top-1 agreement with PyTorch
• bundled Chinese audio: 100% frame top-1 agreement and exact transcript 开饭时间早上九点至下午五点

The generated plan and model weights are intentionally not committed because TensorRT plans are tied to the target GPU architecture and TensorRT version.

Closes #3453