Use memory pool for memory allocation when sampling && feature collecting

[eedalong]

We create a tensor through pytorch pytorch already use CudaCachingAllocator for memory allocation on device. I also implemented a simple test to check tensor's memory allocator and it shows that a cuda tensor indeed uses cudaCachingAllocator for memory allocation.

[lgarithm]

What about the performance of cudaCachingAllocator?

I didnt measure it, but I think it's a pretty good memory pool. but I'll do more research on this CudaCachingAllocator.

[eedalong]

We did some experiments on cudaCachingAllocator and it shows great performance. So we will close this issue.