Performance improvements in generation of FESpaces

Some improvements in the performance of FESpaces are needed, in particular:

- [x] Improve efficiency of low level routines that generate dofs, do the interpolation etc (i.e. introduce function barriers and avoid fine grained allocations). **Fixed since PR #54**
- [ ] Improve efficiency of the interpolation on Dirichlet boundary. (i.e., only loop over boundary cells, some times several tags can have the same function associated to them)