Fixes https://github.com/Itseez/opencv_contrib/issues/332

Relevant material:
http://stackoverflow.com/questions/6066273/pointer-into-vector-but-not-into-a-valarray
http://objectmix.com/c/40001-address-first-member-valarray.html
http://www.open-std.org/jtc1/sc22/wg21/docs/lwg-defects.html#389

I have to admit my naivety here, this issue is was complex for me
to diagnose. It appears that in the MSVC 2008/2010 implementation
of std::valarray, when indexing and assigning to a const variable,
that the const version of the member function
valarray<T>::operator[](size_t) returns a value instead of a
const-reference. Therefore, taking the address of this is a
compiler error as you can't take the address of an lvalue. By
performing a const_cast on the array we change which function
is called and thus we actually do recieve a const-reference.
This should not effect other compilers.

Adding float casts allows them to compile on MSVC and should not
affect other compilers.

NAN is a compiler specific constant that does not exist on
visual studio. I also noticed that some NAN instances had been
cast to float, whereas quiet_NaN is both 1) defined by the
standard and 2) templated for the correct type. This is a much
more portable method of getting a NaN value.

Just adding a float cast (since a float is being assigned to)
allows MSVC to compile and should not affect other compilers.

round does not exist in MSVC<2012 - added with the same code
as in line_descritor/binary_descriptor.cpp

Adding the decimal point means that VS doesn't complain about
lack of overrides for sqrt with integers.

vector.data() does not exist in MSVC<2012, so I changed it to
&vector.front() which has the same effect (pointer to the first
element in the vector).

Adds example on segmented word recognition. Shows the use of the OCRHMMDecoder with the NM and CNN default classifiers.

Overload the run() method in BaseOCR class in order to adapt to different classifier callbacks. The original run() method accepts only one Mat input image, this is expected to be a binarzed image with black and white text and works both with the OCRTesseract class and the OCRHMMDecoder class when the character classifier callback works with binary images (e.g. NM). The new run() method accepts two Mat input parameters. One for the gray scale (or color) source image and the other for a binary mask where each connected component corresponds to a pre-segmented character in the input image. This way the OCRHMMDecoder is able to work with character classifiers that operate in grey scale (or color) images (e.g. a CNN).

Better CNN model for character recognition. Trained with an augmented dataset by adding translation/scale variations. Updated the croped word recognition with new class numbering (compatible with previous NM classifier).

Now the filter natively supports StereoBM and StereoSGBM with no
parameter tuning required. Also, now user won't need to set the ROI and
the right matcher parameters manually, it is all done in the respective
convenience factory method based on the left matcher instance. Tutorial
was added to clarify the provided example of use.

Fix bug #4373: 

Fix memory leak bug #4420

Fix bug #4373: Error (Assertion failed in resize) when passing very elongated contours to the recognition module

Fixed bug in LSDDetector where mask doesn't remove all undesired lines

update to use the new createOCRHMMTransitionsTable() function, and fix program description in header comments

Adds createOCRHMMTransitionsTable() utility function to create a tailored language model transitions table from a given list of words (lexicon)

TLD Open CL Support

wrap DAISY and LATCH to scripting