Get base pointer from Allocation

[From00]

PR types

New features

PR changes

Others

Describe

For performance considering, Paddle caches and manages the memory malloced from devices. The memory pointer given to user is not the base-pointer that directly returned from the device interface, as it may have undergone many transformations such as split and align by Allocator.
However, sometimes the base-pointer is required. For example, cudaIpcGetMemHandle needs a base-pointer of the device allocation created with cudaMalloc, so it can achieve memory sharing between different processes.
This PR maintains the base-pointer information in all allocators existing in Paddle, and implement a base_ptr interface in Allocation class. Thereby we can get the base-pointer from an allocation like:

paddle::platform::CUDAPlace& place = paddle::platform::CUDAPlace(); 
size_t size = 1024; // bytes    
std::shared_ptr<Allocation> allocation = paddle::memory::AllocShared(place, size);
void* base_ptr = allocation->base_ptr();

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zhiqiu]

LGTM