Run for module generation in parallel

[wilzbach]

Currently dpl-docs on dlang.org run a single-thread. That makes them the bottleneck in time consumption.

[s-ludwig]

Unfortunately this change causes multiple race conditions. I think when we make an attempt to parallelize document generation, we should also use pure and shared/immutable within the affected call tree to let the compiler make reasonably sure that the approach is correct. Frankly, I don't feel comfortable to do multi-threading these days without either that or guaranteed memory isolation, unless it is a small and isolated piece of code. Too many bad memories from C++ times ;-)