Nvidia

We should add an example that demonstrates how to use a CUDA stream with thrust algorithms, e.g:

thrust::reduce(thrust::device.on(stream), ...);

Describe the bug
If min_samples_split is a small float, then it may be equivalent to splitting on < 2 samples. This causes cuml to blow up:

RuntimeError: exception occured! file=../src/decisiontree/decisiontree.cu line=41: Invalid value for min_samples_split: 1. Should be >= 2.
Obtained 64 stack frames
#0 in /home/mboling/miniconda3/lib/python3.8/site-packages/cuml/common/../../..

Describe the solution you'd like

Improve this error to provide both the found hardware and the expected hardware

 No compatible device was found for instantiating TensorRT **engine**

Ref: https://github.com/NVIDIA/cub/blob/main/CONTRIBUTING.md#development-model