RuntimeError: CUDA error: invalid argument when running `tracer.export_caffe2()`

## Instructions To Reproduce the Issue:

Run my ['Detectron2: Train and export model.ipynb'](https://colab.research.google.com/github/TannerGilbert/Object-Detection-and-Image-Segmentation-with-Detectron2/blob/master/Detectron2_train_and_export_model.ipynb#scrollTo=3O22vkBFj_6u) notebook.

## Expected behavior:

Detectron2 should let me export my object detector as a caffe2 model as described in the [export documentation](https://detectron2.readthedocs.io/en/latest/modules/export.html), but the following error occurs when running [`export_caffe2()`](https://detectron2.readthedocs.io/en/latest/modules/export.html?highlight=export_caffe2#detectron2.export.Caffe2Tracer.export_caffe2).

Error:
```
RuntimeError: CUDA error: invalid argument
CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
```

## Environment:

Google Colab

Run the linked Detectron2 training and export notebook in Google Colab, then reproduce the failure at tracer.export_caffe2(). Read the export_caffe2 documentation and rerun with CUDA_LAUNCH_BLOCKING=1 to obtain a synchronous traceback. Done means the object detector exports as a Caffe2 model without the CUDA invalid-argument error.

Written by the indexing model from the issue text.