update gradient operator registry mechanism

Implement distributed training save model, improve master.NewClient i…

do not do log.Errorln when checkpoint is not found (which is normal)

OLD: op_type -> grad_op_creator

NEW: grad_op_type -> grad_op_creator
     op_type -> grad_op_type

Backward

Fix doc

Replace `OperatorPtr` with `std::shared_ptr<OperatorBase>`

Remove ScopePtr and OperatorPtr

* ScopePtr means pointer of scope, but it can be shared or uniqued.
Change it to std::shared_ptr<Scope> to make code better to read.

* Expose Net to Python

* Expose PlainNet to Python, make python can add_op, complete_add_op
* Provide a low level api to manipulate Net
* Unittest for Net::DebugString

Fix memory deps on device context

Move Copy out from memory.h into memcpy.h

Slight simplification of eigen.h in hope to make it more readable.

Add CPU/GPU Memcpy in memory folder

Simplify Tensor implementation