Add proper linear discriminant analysis for supervised feature selection

[jvoigts]

The current method is ok but doesnt do very well in some cases if there's significant covariance between waveform samples across classes.