Parallel rustc spends a lot of time creating threads

[alexcrichton]

After some discussion on zulip it looks like rayon will immediately spawn all threads for the thread pool on startup, but this isn't neessarily suitable for rustc's use case. In a profile I captured of compiling Cargo the first set of crates being compiled spent a huge amount of time just spawning threads.

I ran perf at the start of cargo build and killed it after a second or so, and the profile output of this (cleaned up) aggregated across all builds was:

I think that roughly means that 80% of the cpu time was spent transitively spawning threads/processes near the start of the build, and that seems a bit excessive!

cc @Mark-Simulacrum @Zoxc