normal bayes classifier has been parallelized using TBB; letter_recog sample updated to demosntrate knearest & bayes classifiers (thanks to Konstantin Krivakin for the patches)

* if the filename looks like a URI, it is opened in non-blocking mode, cvQueryFrame() could skip frames or grab one frame more than once
* if the filename looks like a filename, it is opened in blocking mode. cvQueryFrame() grabs consecutive frames
* otherwise the filename is interpreted as a gstreamer pipeline as used with gst-launch. The last element of the pipeline has to have the property name=to-opencv

added downsample function into gpu module, refactored it a little bit, added guard for CUDA related include in cascadeclassifier_nvidia_api.cpp

[~] Refactored, cleaned up, and consolidated the code of GPU examples (cascadeclassifier and cascadeclassifier_nvidia_api)

improved tree_engine.cpp sample (added train file data specification; print sorted variable importance table)

[*] Fixed #974 ("GPU CascadeClassifier fails with some training files"): Moved IsNodeLeaf bit from NodeDescriptor to FeatureDescriptor for both left and right nodes, therefore from now on max number of rects in a feature is 31