Optimize distributed training performance

[typhoonzero]

No description provided.

[wangkuiyi]

It is difficult to understand the title of this issue.

[typhoonzero]

Sorry, the original issue is trying to leave a TODO in the project. I'm writing a transpiler which can split one variable and send to multiple parameter servers to do the partial, elementwise optimization, but the current implementation of split_op have to copy the splited data to newly allocated memory, which can be optimized.

Then we noticed that all the allocation is managed by "buddy allocator" under paddle/memory, if we use several pointers pointing to sections of the original tensor, and then use these pointers in new tensor objects will probably cause memory errors.

I changed this issue to be more general, and we can do some other performance enhancements for now.