The parallel upsweep needs to know about _all_ the modules that make up
a module loop in order to correctly determine a module's external loop
dependencies. Otherwise, incorrect dependency information may be
computed, resulting in an eventual deadlock during compilation.

So don't filter boot modules in getModLoop, and instead have its callers
filter them when needed.

Following this change, GHC could compile itself via --make -O2 -j. Yay!