added opimized versions of LRP for linear layers for simple and eps.
added an "lrp aware" forward pass, reducing total run time when for the same X multiple relevance maps should be computed (e.g. with different target classes or param settings) at the cost of a slower forward pass