minor refactoring of gpu tests

#1713 Added the possibility of setting user_block_size manually for gpu::matchTemplate function (and gpu::convolve). Added a buffer param into these functions. Removed using of 2^n block sizes when it's not necessary.

used random images in gpu filter tests

added assertion on double types for old devices

added additional tests for gpu filters
fixed gpu features2D tests

fixed some bugs in gpu module

refactored tests for them

fixed several warnings

wrote more complicated tests for them
implemented own version of warpAffine and warpPerspective for different border interpolation types
refactored some gpu tests

updated gpu tests for CornerHarris and CornerMinEigen
moved direct convolution implementation to gpu::filter2D, gpu::convolve now use only DFT-based algorithm (Bug #1639)

removed void integral(const GpuMat& src, GpuMat& sum, GpuMat& sqsum, Stream& stream) - it fails with NPP_NOT_IMPLEMENTED error
updated docs, accuracy and performance tests

now it supports kernel's size up to 32

added PyrLKOpticalFlowDense performance test

fix createOpticalFlowNeedleMap

update gpu tests