* fix gpu memory allocation policy

* refine codes

* fix code style

* follow comments

* add gflags for C++ unittest

* make enforce a target and dependent on nccl when gpu is enabled

* add some more dependency

Use posix_memalign to allocate aligned memory, since some SIMD instructions require the alignment of memory accesses.

* Add debug logs in scope, meta_cache and memory

* Add missing deps

* Init

* Stash

* Polish SaveLoadOp

* Fix CI

* Polish code

* Save GPU Tensor

* Stash

* Fix CI

* Make global scope not thread-safe

1. It is no need to make global scope thread-safe, since it will be
invoked in Python main thread.
2. Do not free the global scope when C++ exit. Let the OS free memories,
otherwise, we need to handle the destroy dependencies.

See
https://google.github.io/styleguide/cppguide.html#Static_and_Global_Variables

* Revert "FIX: Release CPU/GPU memory via deleter"

This reverts commit 8f80f5bc.

By shell command

```bash
sed -i 's#ifdef PADDLE_ONLY_CPU#ifndef PADDLE_WITH_GPU#g' `find ./paddle/ -name '*.h' -o -name '*.cc' -o -name '*.cpp' -o -name '*.c' -o -name '*.cu'`
sed -i 's#ifndef PADDLE_ONLY_CPU#ifdef PADDLE_WITH_GPU#g' `find ./paddle/ -name '*.h' -o -name '*.cc' -o -name '*.cpp' -o -name '*.c' -o -name '*.cu'`
```

User StridedMemCpy in Concat/Split Op

* with unit-tests
* Also complete `memcpy`