Perf tweaks for many entities

Small changes to remove perf bottlenecks with many (1e4) entities:

• avoid creation of Domain() (return const ref instead)
• move getEntities and getElementTypes to pre-processing phase (they both do a find in an std::set)
• remove test on element type in assembly loop on entities

@ARoyer : have a look at let me know, especially for the last one (I've left the original code inside an #if 0).