* Make StreamSafeCUDAAllocator compatible with NaiveBestFit strategy

* Set FLAGS_use_stream_safe_cuda_allocator to false

* Update

* Remove unnecessary code

* Fix CI errors

* Add UT