Most codes are ported from AMD's Bolt library.
Four methods are implemented:

SORT_BITONIC,   // only support power-of-2 buffer size
SORT_SELECTION, // cannot sort duplicate keys
SORT_MERGE,
SORT_RADIX      // only support signed int/float keys

Test cases (accuracy and performance) are provided.

The bug is a buffer overrun when border type is reflect101.
It is found that gfft crashed with input of size 100x100 on Intel CPU.

In current implementation, this function only works when anchor point is
in the kernel center and kernel size supported is either 3x3 or 5x5.

For example, haarcascade_frontalface_alt2.xml is now supported.
Note that classifier's pattern of a cascade file must be consistent,
i.e., all trees must either have two nodes or one node, otherwise
unexpected results will occur.

Other fixes:
Test cases are updated.
Some unused codes are removed.
Fix some problems of haar when using OclCascadeClassifierBuf.

This fix is a workaround for current 2.4 branch without introducing an
additional oclMat buffer into CannyBuf object.
Test case is cleaned up.
Volatile keywords in kernels are removed for performance concern.

Additional notes with this commit:
1. Add cornerHarris_dxdy and cornerMinEigenVal_dxdy to get
the interim dx and dy output of Sobel operator;
2. Add minMax_buf to allow user to reuse buffers in minMax;
3. Fix an error when either min or max pointer fed into minMax is NULL;
4. Corner sorter temporarily uses C++ STL's quick sort. A parallel
 selection sort in OpneCL is contained in the implementation but disabled
due to poor performance at the moment.
5. Accuracy test for ocl gfft.

Use predefined OpenCL function to convert integers to floating points.
This is more accurate than before as it enables:
1. saturate cast
2. customized rounding

Speedup about 30% on 6730M GPU.

1. Use macro defines for some parameters(radius).
2. Reduce local memory usage.
3. Fix accuracy problem on Intel GPU.