fixed BFM warning ("cannot tell what pointer points to")

added assertion after all kernels calls

created element_operations and matrix_reductions files for farther refactoring

refactored transpose_gpu, made it non template function.

module reorganization: added folder with pure device functions, cuda_shared.hpp renamed to internal_shared.hpp

refactored gpu module.

  fixed some bugs in init_data_cost on first level (added non-reduction version for first level)
  optimized compute_data_cost like init_data_cost (used reduction scheme)
  avoid temp matrix

some refactoring of StereoBeliefPropagation.