Not sure this is a clear win. Some notes:

This uses the simulation from the unit tests as a representative
workload.  This workload might be suboptimal in some ways:

- Quite slow, 55 seconds for precompilation on my laptop.

- Uses ficticious Wolpertinger species instead of Skylark, this might
  not be optimal for the real simulation.

- There might be unnecessary data stored in the precompile file, this
  could be reduced with `@setup_workload` from PrecompileTools.

Other notes:

- Need to set up a representative benchmark, measuring changes in
  package load time and TTFX.

- After running precompilation from the Pkg REPL, and then running
  tests, precompilation runs again (taking 55 seconds again).  Why?