Vectorization with a cost model

There is many more things to be done here:

• The flopcost model from permutation must be flop-exact
• The permutation must be taken into account when determining flop cost of a kernel
• Hardware parameters (peak performance, memory bandwidth) must be read through ini file
• Collect ideas for ILP heuristic
• Do not store quadrature point size with the tabulations - it may be altered globally
• write generator for all vectorization opportunities
• Rewrite old vectorization strategies in terms of a cost function
• Do meaningful stringification of vectorization strategy dictionaries
• Rethink options interface for vectorization