Vectorization strategy selection through Autotuning

This is the MR for #125 (closed)

Tasks:

• Generate a standalone C++ file for benchmarking a given sum factorization kernel
• Above with fastdg
• Above with h/l
• Implement a way to do JIT compilation of benchmark programs (codepy?)
• Asynchronize the cost function evaluation (instead of using min)
• Add a pickle-cache for benchmarking results
• Add an autotune cost model implementation
• Think of an interface how benchmark programs communicate their measurements to the code generator
• Tests