Counterintuitive comparison of matrices with NaN values (Bug #3011)

At the moment element-wise comparison of matrices works in OpenCV as follows:

• the comparison operations A < B (and A > B) are calculated element-wise
• but the comparison operations A <= B (and A >= B) are calculated as the logical negation of the compariosn A > B (correspondingly, A <B)

(see the code of the function cv::compare in the file source:modules/core/src/arithm.cpp and the template function cmp_ in the same file)

Since in C++ the operations <, >, <=, and >= with NaN values always return false, there is the following counterintuitive behavior:
if a matrix A contains only NaN values, we will have

countNonZero(A < 0.5) ==> 0;
countNonZero(A > 0.5) ==> 0;
countNonZero(A <= 0.5) ==> A.total();
countNonZero(A >= 0.5) ==> A.total();

This may be fixed if the comparison functions will make element-wise calculations for non-strict inequalities too. But in this case all the optimizations (SSE, etc) should be changed too.