* improve `CUDA_TARGET_CPU_ARCH` cache initialization,
    allow to override initial value from calling script;
* add `CUDA_TARGET_OS_VARIANT` option to select OS variant;
* add `CUDA_TARGET_TRIPLET` option to select target triplet from
    `${CUDA_TOOLKIT_ROOT_DIR}/targets` folder;
* remove `CUDA_TOOLKIT_TARGET_DIR` option, now it is calculated from
    `CUDA_TARGET_TRIPLET`, the old approach still can be used for compatibility;
* for CUDA 6.5 and newer try to locate static libraries too, because
   in 6.5 toolkit for ARM cross compilation only static libraries are included.

move main CUDA group to modules/core/cuda.hpp

add a note to use new cudev module as a replacement

it is internal class, no need to export it

The deinitialization of BufferPool internal objects is controled by global
object, but it depends on other global objects, which leads to errors
caused by undefined deinitialization order of global objects.

I merge global objects initialization into single class, which performs
initialization and deinitialization in correct order.

it is generated by CUDA headers and we can't fix it

to compensate for neon ieee754 non-compliancy.
Also changed the comparison between max valid and calculated distance to
make the error message more accurate (in case curMaxDist == maxDist)

Set it 1 instead of 0.001, as is already done in gaussianBlur3x3. That
will allow integer destination matrices that are not exactly the same,
but very close to the expected result, to pass the test.

NEON speedup: 2.31x
Auto-vect speedup: 2.26x

Test kernel: [-0.9432, -1.1528, 0, 1.1528, 0.9432]

NEON speedup: 2.36x
Auto-vect speedup: 2.36x

Test kernel: [0.1, 0.2408, 0.3184, 0.2408, 0.1]

NEON speedup: 9.46x
Auto-vect speedup: 1x

Test kernel: [-0.9432, -1.1528, 0, 1.1528, 0.9432]

NEON speedup: 8.64x
Auto-vect speedup: 1x

Test kernel: [0.1, 0.2408, 0.3184, 0.2408, 0.1]

NEON speedup: 2.12x
Auto-vect speedup: 1.01x

Test kernel: [-2, 0, 2]