* Moved common code required by CUDA and OpenCL caching algorithm into
  kernel_cache.[hpp|cpp]

* Added common/compile_kernel.hpp header that defines the signature
  of the function, compileKernel, that each backend has to implement.

* Each backend has to implement/satisfy the following requirements:

  - Provide compile_kernel.cpp source with compileKernel function that is
    used by common::findKernel

  - Provide Kernel.hpp/cpp that implements KernelInterface from
    common/KernelInterface.hpp

  - Kernel.hpp also provides a functor than helps launch backend kernels.

* Moved kernel utility helpers into separate header(s)/source:

  - TemplateArg.hpp/cpp contains the TemplateArg struct and some helper macros
    to convert template arguments to strings.

  - TemplateTypename.hpp contains the templated TemplateTypename struct that
    helps to convert backend kernel paramters to TemplateArg object.

* Refactored all CUDA kernels to use the new caching API

* Refactored only transpose, morph and canny to use new caching API from OpenCL

* Reduced lot of unnecessary instantiations for morphological functions