Learning backend cuda

Context

This MR is intended to add the forward implementations for Accuracy computation operators:

• And
• ArgMax
• ReduceSum

And the forward implementations for operators used in the MSE #20 (moved):

• Mul
• ReduceMean
• Pow

Also a fix to #23 (moved) is added: update Tensor descriptor when shape changes.
Also a fix to #7 (moved) is added: force conv fwd algorithm.

We also add operator Ln forward implementation to support #24 (moved).

Modified files

• AndImpl.hpp, AndImpl_CUDA_kernels.hpp, AndImpl_CUDA_kernels.cu, AndImpl.cpp and Tests_AndImpl.cpp, add And forward impl;
• ArgMaxImpl.hpp, ArgMaxImpl_CUDA_kernels.hpp, ArgMaxImpl_CUDA_kernels.cu, ArgMaxImpl.cpp and Tests_ArgMaxImpl.cpp, add ArgMaxforward impl;
• ReduceSumImpl.hpp, ReduceSumImpl_CUDA_kernels.hpp, ReduceSumImpl_CUDA_kernels.cu, ReduceSumImpl.cpp and Tests_ReduceSumImpl.cpp, add ReduceSum forward impl;
• LnImpl.hpp, LnImpl_CUDA_kernels.hpp, LnImpl_CUDA_kernels.cu, LnImpl.cpp and Tests_LnImpl.cpp, add Ln forward impl;
• ConvImpl.hpp, ConvImpl.cpp and Tests_ConvImpl.cpp, force Conv Fwd algorithm;
• TensorImpl.hpp, update Tensor descriptor when tensor shape has changed;
• MulImpl.hpp, MulImpl.cpp and Test_MulImpl.cpp, add Mul forward() and backward();
• ReduceMeanImpl.hpp, ReduceMeanImpl.cpp and Test_ReduceMeanImpl.cpp, add ReduceMean forward() and backward();
• PowImpl.hpp, PowImpl.cpp and Test_PowImpl.cpp, add Pow forward() and backward();

TODO

• Mul
• ReduceMean
• Pow
• And
• ArgMax
• ReduceSum
• Fix TensorImpl