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

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)

replaced alloca() (a.k.a. cvStackAlloc) with AutoBuffer or vector() everywhere. cvStackAlloc() is still defined, but we do not need alloca() anymore to compile and run OpenCV (fixes #889 and may be some others)