Adapt a ufunc-like architecture for calls

[Ivorforce]

This way, we should be able to

• better dispatch for different SIMD support, see Add support for SIMD multi-arch exports #102
• better define and understand supported types from a meta perspective
• Simplify the c++ code generation logic
• Reduce the binary size by having even fewer branches

We will probably have to generate tables from python. Either way it's good to have the information statically.
We will also need to abstract broadcasting first (#162).