CUDA error 400: invalid resource handle

Error encounting when working on Decord + TensorRT for inference time.

## Reproduce
Code:
```python
def _test2():
# for tensorrt stuff
import pycuda.autoinit
import decord
VIDEO_PATH = 'data/test_video.mp4'

vr = decord.VideoReader(VIDEO_PATH, ctx=decord.gpu(0))
frames = vr.get_batch([0,1,2])
print(frames.shape)

if __name__ == '__main__':
_test2()
```

Output with error:
```
CUDA error 400 at line 34 in file /workspace/libs/decord/src/video/nvcodec/cuda_mapped_frame.cc: invalid resource handle
Terminated
```

Remove the line `import pycuda.autoinit`, code exec normally without any error.
```python
def _test2():
# for tensorrt stuff
# import pycuda.autoinit
import decord
VIDEO_PATH = 'data/test_video.mp4'

vr = decord.VideoReader(VIDEO_PATH, ctx=decord.gpu(0))
frames = vr.get_batch([0,1,2])
print(frames.shape)

if __name__ == '__main__':
_test2()
```

This may relate to [this issue](https://github.com/NVIDIA/DALI/issues/408)

**Keywords**: TensorRT, PyCUDA