added support of CV_TM_CCORR (via FFT) into gpu::matchTemplate (versions both with block and without blocks)

refactored transpose_gpu, made it non template function.

"descriptor-opponent-sift",       //ticket 737
"_3d-rodrigues",                  //ticket 435

module reorganization: added folder with pure device functions, cuda_shared.hpp renamed to internal_shared.hpp

added first version of gpu::matchTemplate, currently it works only with 8UC1 images and SQDIFF method

revert to using fixed RNG in EM, otherwise the algorithm becomes non-deterministic and the test fails sometimes. in the meantime, EM needs to switch to more robust cv::kmeans instead of CvEM::kmeans.

turn off SSE2 optimization in cv::exp & cv::log in the case of GCC <=4.1 (not no Mac, where a patched GCC 4.0 is used)

fixed compile error in transform.

turn off SSE2 optimization in cv::exp & cv::log in the case of VS2005 or earlier versions, because of missing _mm_cast* intrinsics (ticket #729)

fixed several typos in docs; make MLData capable of reading csv files with much more columns than before