Created by: zhiqiu

PR types

Function optimization

PR changes

Others

Describe

Refine best_fit_allocator_test to support cuda-10.2

We found that the CUDA memory usage of each CUDADeviceContext increases from CUDA-10.1 to CUDA-10.2.

The original implementation of best_fit_allocator_test creates a new CUDADeviceContext for each thread and a large number of contexts may use off CUDA memory.

This PR fixes that problem.