Support compiling codegen units in parallel

[bjorn3]

With this on machines with many cores cg_llvm no longer has a perf advantage over cg_clif on the simple-raytracer benchmark:

$ hyperfine --prepare "cargo clean" "cargo build" "../build/cargo-clif build"
Benchmark 1: cargo build
  Time (mean ± σ):      6.782 s ±  0.014 s    [User: 23.918 s, System: 4.269 s]
  Range (min … max):    6.755 s …  6.800 s    10 runs
 
Benchmark 2: ../build/cargo-clif build
  Time (mean ± σ):      6.505 s ±  0.017 s    [User: 16.604 s, System: 3.640 s]
  Range (min … max):    6.484 s …  6.540 s    10 runs
 
Summary
  '../build/cargo-clif build' ran
    1.04 ± 0.00 times faster than 'cargo build'

And even on my personal laptop there is a non-trivial perf improvement (~35% perf benefit of cg_clif over cg_llvm as opposed to ~20% previously):

Benchmark 1: cargo build
  Time (mean ± σ):     16.985 s ±  0.487 s    [User: 48.250 s, System: 4.623 s]
  Range (min … max):   16.692 s … 18.339 s    10 runs
 
Benchmark 2: ~/Projects/cg_clif2/build/cargo-clif build
  Time (mean ± σ):     12.517 s ±  0.065 s    [User: 29.892 s, System: 4.536 s]
  Range (min … max):   12.438 s … 12.642 s    10 runs
 
Summary
  '~/Projects/cg_clif2/build/cargo-clif build' ran
    1.36 ± 0.04 times faster than 'cargo build'

On the rustc perf suite there are now only a couple of wall time regressions compared to cg_llvm as opposed to almost all.

Wall time on the rustc perf suite when compared to cg_llvm.

The jobserver handling could probably be finetuned for a bit more performance, but as I have been struggling with getting it to work correctly at all for almost a week, I'm going to leave that for now.

Fixes #652

See #1266 for the previous PR in this series.