Remove some deepcop[ies] to speed up strengthReduction

[davschneller]

This PR does:

• removes all deepcopy calls in strengthReduction which were used to support GPU caching for fused GEMMs—by cloning the cost estimator class which unfortunately also meant to copy the whole _cache from the bounding box cost estimator. Instead, we now save a caching set per node, not per cost estimator class.
• adds some generator expressions, removes some temporary arrays

[krenzland]

Looks mostly good, one small question.