PCA with retained variance on double (Bug #3068)

Hello,

I came across a bug in cv::PCA when setting the amount of variance to be retained.

The bug is in matmul.cpp on line 2975 and downwards (https://github.com/Itseez/opencv/blob/master/modules/core/src/matmul.cpp)

The matrix g is of type ctype, but g.at uses the float template. If the input matrix is of type double, ctype is CV_64FC1, so .at<float> will give back the wrong values. Furthermore I am not sure why it doesn't use something like this :

@Mat g = eigenvalues / cv::sum(eigenvalues)[0];
double sum = 0.0;
int L;
for (L = 0; L < g.rows; L++) {
sum += g.at<float>(L,0);
if (sum >= retainedVariance)
break;
}@

But that is perhaps another matter. ps. this has the same .at<float> bug, but I am not sure how this should be done generally :)