Investigate JIT throughput for huge grammars

[bbannier]

It seems that for huge grammars (hundreds of units, thousands of LOC) JIT throughput can be very slow. We should benchmark such grammars to make sure we are not behaving worse than absolutely necessary.

Things which would be interesting to look at:

• does the resolver time scale at most linear with the number of names/types?
• does the number of template instantiations scale linear with the number of types?