A small simplification exercise showing that for all processes the eps_prev can be removed w/o loss of performance. Same could be done with eps, but requires refactoring of secondary variables extrapolation.
Tested on LARGE ctest hollow-sphere. No measurable cpu-time difference. Memory is saved, there are 73.2MB less data allocated ~64bytes per integration point, which could be sizeof(eps_prev) x sizeof(double) + padding? = 6x8 + 24?
The only problem appeared for the back-filling (subdomain deactivation) case where a different strain was stored. This is a not so simple issue to be dealt with in the future.