feat: enable Jetson TensorRT CLI profiling

[hyeokjun32]

Summary

• enable TensorRT runtime execution path on Jetson
• add TensorRT runtime buffer allocation and device pointer preparation
• add CUDA driver/context setup for TensorRT device allocation
• add TensorRT dummy input generation and inference execution support
• add Python 3.10 compatible TOML loading fallback
• validate Jetson CLI profiling flow for TensorRT engines

Test

• poetry run python -m py_compile edgebench/engines/tensorrt.py
• poetry run python -m py_compile edgebench/config.py
• poetry run python -c "from edgebench.engines.tensorrt import TensorRtEngine; print('TensorRtEngine import ok')"
• poetry run python - <<'PY'
  from edgebench.config import resolve_compare_thresholds
  print(resolve_compare_thresholds())
  PY
• Jetson: python3 - <<'PY'
  from edgebench.engines.tensorrt import TensorRtEngine
  engine = TensorRtEngine()
  engine.load("models/resnet18.onnx", engine_path="models/resnet18.engine")
  feeds = engine.make_dummy_inputs()
  outputs = engine.run(feeds)
  print(len(outputs))
  PY
• Jetson: python3 - <<'PY'
  from edgebench.engines.tensorrt import TensorRtEngine
  engine = TensorRtEngine()
  engine.load("models/yolov8n.onnx", engine_path="models/yolov8n.engine")
  feeds = engine.make_dummy_inputs()
  outputs = engine.run(feeds)
  print(len(outputs))
  PY
• Jetson: python -m edgebench.cli profile models/resnet18.onnx --engine tensorrt --engine-path models/resnet18.engine --precision fp16
• Jetson: python -m edgebench.cli profile models/yolov8n.onnx --engine tensorrt --engine-path models/yolov8n.engine --precision fp16