* rewritten the following layers to be [much] more efficient: convolution, fully connected, activations (relu, tanh, ...), LRN. Use optional AVX optimization for the first two.

* eliminated trailing whitespaces

* Reuse deep learning output blobs

* Changed order for iterating through blobs while seeking memory. Refactored a little.

Made separate functions for computing output shapes for all layers. Removed output blobs allocation from layers

* Fixed bugs after dnn refactoring. Added new tests for layers

* Fixed torch tests

* the first commit in the merged dnn: convert some public API from Blob's to Mat's

* temporarily or permantently removed OpenCL optimizations, which are not always stable nor usually very efficient; we'll likely use Halide instead

* got rid of Blob and BlobShape completely; use cv::Mat and std::vector<int> instead

* fixed a few compile errors

* got rid of separate .hpp files with layer declarations; instead, put everything into the respective .cpp files

* normalized all the layers' constructors; we concentrate on loading deep networks layers from files instead of constructing them from scratch, so we retained only SomeLayer::SomeLayer(const LayerParams& params); constructors

* fixed sample compilation

* suppress doxygen warnings

* trying to fix python bindings generation for DNN module

* temporarily disable python bindings while we refactor the module

* fix win32/win64 compile errors; remove trailing whitespaces

* fix win32/win64 compile errors; remove trailing whitespaces

Getting layer inputs (#1006)

They are initialized only in functors, not in inherited fields.

moved BLAS/LAPACK detection scripts to the main repository. Give preference to the external BLAS over OpenCV's gemm, which improves performance of dnn module by a lot

fixed some compile warnings in dnn & protobuf; improved convolution layer performance when blas is not available by parallelizing gemmCPU() function in dnn