weights array is only allocated for (l_count+1) elements,
but then weights[l_count+1] element is accessed.

This should allow many algorithms to take advantage of more parallelization
technologies.

This allows subclasses of CvBoost to override initial weights that
update_weights uses without duplicating its entire implementation.

add hack to disable optimization of linear svms; improved precision of optimize_linear_svm; add the relevant test, which however requires some big database (so it's disabled by default)

Initialize local variables

Also corrected some indexes to be size_t.

(Sometimes sample_data became bad because the buffer with its values was deallocated early).

This reverts commit 4d36be87.

Conflicts:
	modules/ml/include/opencv2/ml/ml.hpp
	modules/ml/src/data.cpp

possibly fixed crash boost caused by the recent modification; slightly improved mlp params in letter_recog sample

fixed 2 build problems on MacOSX (highgui with ffmpeg, opencv_world); added macro CV_INIT_ALGORITHM for simpler algorithm registration (ticket #1876)

added dedicated <modname>_init.cpp files with initModule_<modname>() functions and all the relevant structures; made BackgroundSubtractorMOG/MOG2 derivatives from Algorithm; cleaned up MOG2 implementation and made it multi-threaded.