This version uses the revised math for SRW (with M&dM, p&dp, and gradient steps).
Training is much faster and duplicates less computation. Inheritance among the SRW classes is cleaner and requires fewer lines of code for overrides. Regularization is slightly simpler.
2.1 runnables are the same as 2.0 and include Grounder, Trainer, QueryAnswerer, and GradientFinder.