The original motivation for this was trying to get faster turnaround
when running check-llvm at -O0. The impact on debuggability seems
to outweight the benefits, as it makes StringRef and other ADT
classes fairly undebuggable.
Add a section in the developer's documentation explaining that
optimized builds are needed for decent test performances instead.
Based on a discussion on llvm-dev, pending more general consensus.
Should we rephrase this in a way that makes it clear that you don't always have to set the CMake flag, e.g. in release? Maybe something like `build ... in release or at least with -O2"?