Test accuracy for PR#3829.Fix warning on gcc, comment out INIT_RANDOM_SEED for same result each ran.

Current implementation of miniflann is releasing the trained index, and
rebuilding the index from the beginning.
But, some indexing algorithms like the LSH are able to add the indexing
data after that.
This branch is implementation of that optimization for LshIndex
FlannBasedMatcher in the feature2d.

previous implementation was not thread/stream safe, since it used constant
memory

new implementation doesn't use any global objects, so it is thread/stream safe

Add GPU models in seperate folder, only old models are supported and they were replaced to make the CPU interface work.
This solves the undesired consequence of breaking the GPU interface.

move hog::set_up_constants from constructor to compute method

if user changed CUDA device between constructor and computation,
some variables were uninitialized

The architecture has been correctly determined earlier in this
file. No need to do it again.