Fix hash(initializer_list) and systematics improvements

I found the main culprit for slowing down HH DNN outputs with systematics, but in the process of debugging that also a few points where the automatic systematics code could be made more efficient, the main one already included (only branching for the variations that really need it, if a group is mixed between weights-only and kinematic variations - notable example: b-tag SF weight variations and jet/met kinematics); the rest are python speed improvements.