There might be memory leak in instanceNormalizationPlugin

[NHZlX]

There are two places that can cause leaks

First, every enqueue calls cudamalloc without releasing

TensorRT/plugin/instanceNormalizationPlugin/instanceNormalizationPlugin.cpp

Line 211 in 572d54f

CHECK_CUDA(cudaMalloc((void**) &_d_bias, n * nchan_bytes));

Second, cudnn_handle will also be created every time

TensorRT/plugin/instanceNormalizationPlugin/instanceNormalizationPlugin.cpp

Line 217 in 572d54f

CHECK_CUDNN(cudnnCreate(&_cudnn_handle));

[rmccorm4]

May be related: https://devtalk.nvidia.com/default/topic/1068998/memory-leak-in-tensorrt-instancenormalization/?offset=2

CC @rajeevsrao

[nietras]

Not sure if this is the same, but I have observed a memory leak in OnnxRuntime using TensorRT as can be seen microsoft/onnxruntime#2773
Although this model uses BatchNormalization and not InstanceNormalization.

[rmccorm4]

Fixed per #315