Build does not feature-detect kFP8

[jonhoo]

Description

Note: this duplicates #922, but adds significantly more detail, so I recommend closing that and using this instead.

The nvinfer1::DataType::kFP8 was added between 8.5.3 and 8.6.0 of TensorRT, but is currently used unconditionally by onnx-tensorrt. As such, the library no longer builds on platforms that are still using TensorRT < 8.6.0, such as NVIDIA JetPack version 5, which is still on 8.5.2. On those platforms, onnx-tensorrt now fails to build with:

[ 86%] Building CXX object _deps/onnx_tensorrt-build/CMakeFiles/nvonnxparser_static.dir/NvOnnxParser.cpp.o
In file included from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/onnx2trt.hpp:10,
                 from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/ImporterContext.hpp:7,
                 from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/ModelImporter.hpp:7,
                 from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/NvOnnxParser.cpp:6:
/nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/TensorOrWeights.hpp: In member function 'std::string onnx2trt::TensorOrWeights::getType() const':
/nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/TensorOrWeights.hpp:118:38: error: 'kFP8' is not a member of 'nvinfer1::DataType'
  118 |             case nvinfer1::DataType::kFP8: return "FP8";
      |                                      ^~~~
In file included from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/onnx2trt_utils.hpp:10,
                 from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/ImporterContext.hpp:8,
                 from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/ModelImporter.hpp:7,
                 from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/NvOnnxParser.cpp:6:
/nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/trt_utils.hpp: In function 'int onnx2trt::getDtypeSize(nvinfer1::DataType)':
/nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/trt_utils.hpp:26:30: error: 'kFP8' is not a member of 'nvinfer1::DataType'
   26 |     case nvinfer1::DataType::kFP8: return 1;
      |                              ^~~~
/nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/trt_utils.hpp: In function 'onnx::TensorProto_DataType onnx2trt::trtDataTypeToONNX(nvinfer1::DataType)':
/nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/trt_utils.hpp:160:30: error: 'kFP8' is not a member of 'nvinfer1::DataType'
  160 |     case nvinfer1::DataType::kFP8: break;
      |                              ^~~~
In file included from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/ImporterContext.hpp:8,
                 from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/ModelImporter.hpp:7,
                 from /nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/NvOnnxParser.cpp:6:
/nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/onnx2trt_utils.hpp: In function 'std::ostream& nvinfer1::operator<<(std::ostream&, const nvinfer1::DataType&)':
/nix/store/skkb1qrwi5rxjf3f4j4cmf4r4cfq9nnl-source/onnx2trt_utils.hpp:78:30: error: 'kFP8' is not a member of 'nvinfer1::DataType'
   78 |     case nvinfer1::DataType::kFP8: return stream << "float8";
      |                              ^~~~
make[2]: *** [_deps/onnx_tensorrt-build/CMakeFiles/nvonnxparser_static.dir/build.make:76: _deps/onnx_tensorrt-build/CMakeFiles/nvonnxparser_static.dir/NvOnnxParser.cpp.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:5213: _deps/onnx_tensorrt-build/CMakeFiles/nvonnxparser_static.dir/all] Error 2

The use of kFP8 should probably be guarded the same way it is in tensorflow itself (tensorflow/tensorflow#60046), namely by utilizing something like

#endif
#if IS_TRT_VERSION_GE(8, 6, 0, 0)
    case nvinfer1::DataType::kFP8:
      return "kFP8";
#endif

Environment

TensorRT Version: 8.5.2
ONNX-TensorRT Version / Branch: 8.6-GA
GPU Type: Jetson AGX Orin
Nvidia Driver Version: JetPack 5.1.2
CUDA Version: 11.4.19
CUDNN Version: 8.6.0
Operating System + Version: JetPack 5.1.2
Python Version (if applicable): 3.12
TensorFlow + TF2ONNX Version (if applicable): 2.16.2

Steps To Reproduce

• Build onnx-tensorrt against tensorrt 8.5.2