Fix dist train broadcasting bug

* 1. Create buddy allocator in each places before NcclBcast the variables
2. Check the memory usage of ALL gpus rather than the first one

* 1. Make NCCLGroupGuard guards only the ncclBcast part, which avoid ncclGroupEnd blocking the exception throwing
2. NOTE the usage of NCCLGroupGuard

* Remove the memory usage check of gpus

* Fix code style

replace use_event with use_cuda, because use_event means the program running with CUDA, so use_cuda maybe more intuitive.

* Use Builder Chain to decorate new builders. It is easy to extend
  builders.
* Make graphviz path as a build strategy, not a FLAGS.

* Eagerly drop local scope in iteration

* Correct create var

* Fix typo

* Debug

* Support testing during training by ParallelExecutor.

* Add unit test.

* Improve the interface.

* Follow comments.