LDA classifier is scale dependent which is wrong

See issue #147 for details. The executive summary is that the classifier should not depend on the scale of the input feature, so an absolute tolerance is wrong. There are probably similar issues with other classifiers (I have not checked).

Fixed by making the tolerance relative to the highest eigenvalue with a "standard" ratio of 1e-5 (this is usual in pseudo-inverse). The other change is just a minor computation improvement which avoids computing a matrix product twice.