Fix: Learning metrics

Context

To Support Adam Optimizer, this MR adds operator Sqrt's forward implementation. It also fixes ReduceMean and ReduceSum in case no reduction is needed (encountered in BCE and MSE unit tests)

Modified files

• SqrtImpl.hpp, SqrtImpl.cpp, SqrtImpl_CUDA_kernels.hpp, SqrtImpl_CUDA_kernels.cu and Test_SqrtImpl.cpp, add Sqrt forward and backward operation;
• ReduceMeanImpl.cpp, fix ReduceMean forward;
• ReduceSumImpl.cpp, fix ReduceSum forward;

TODO

• Add Sqrt
• Fix ReduceMean
• Fix ReduceSum