[WIP] Tensor datatypes, conditionals, and simple nonlinear operations

[davschneller]

Draft due to some work still needing to be completed. This PR adds:

• tensor-level datatypes. (and addressing; even though that might work better on a kernel level).
• very simplistic conditional execution (mainly interesting for the GPU kernels). Not happy with the required CNF/boolean handling implementation; but I guess everything else might need even more refactoring.
• abstract the external API a bit more.
•  add elementwise nonlinear operations, as well as a datatype cast.
• [partially done; the generalization is still missing] Allow matrix multiplications with different precisions, on different rings (e.g. and/xor), and more reduction types; i.e. introduce Elementwise (Op) and Reduction (Op) / Accumulate (Op) operations; both operating on some pre-defined operation. Thus, some of the previous linear operations also become special cases of this: Product -> Elementwise (Mul), ScalarMultiplication -> Elementwise (Mul) (maybe) and IndexSum -> Reduction (Add) as well as Add -> Accumulate (Sum).
• [maybe TODO: add simple emulation routines?]

Besides, refactor all format strings.