Master new signalmodel review

• added more comments to point to Aaron's technical notes.
• This also fixes the negative \sum\lambda^2\eta^2 happening to some IMBH templates.
• Given the new signal model without KDE, in order to avoid nan values returned, the 2D histogram is not globally normalized to the number of samples, n, but pre-normalized in log scale for each SNR column and let normalize() take care of some numerical error to be normalized in the end. We can ignore n as it won't be KDE-ed anyway.

See this review wiki for some discussion about nan issue.