Reduce build/test times

[edwintorok]

This reduces uncached full build times by about 15%, and test times by about 10%.

It does this by:

• avoiding to build bytecode for libraries (though libraries with public_name are still built as bytecode due to a dune bug modes not respected in combination with library with public_name ocaml/dune#3467). This could potentially be further reduced by some of the .opam cleanup/unification that I'm working on.
• using a .mli for large files like server.ml (it only exports a single function, this alone cuts build times by about 15%)
• avoiding to preprocess files that don't use PPX features
• rearranging modules and removing unneeded dependencies such that some of the tests could run earlier, in parallel with the server.ml/client.ml build

Interestingly the koji build times are almost exactly the same, so this only seems to help with dev builds (i.e. dune build and dune build @check), not release buids (release builds introduce additional dependencies and reduce parallelism due to also enabling cross-module inlining).

Some of these commits have been around since 2023, although at the time there were some inexplicable test failures in forkexecd that only happened on the CI (likely exposing latent bugs): #4939
In this PR I've only cherry-picked the build changes, and didn't cherry-pick the forkexecd changes.

---

hyperfine --min-runs 3 'dune clean; dune build --cache=disabled' 'cd ../scm-prev; dune clean; dune build --cache=disabled'
Benchmark 1: dune clean; dune build --cache=disabled
  Time (mean ± σ):     79.936 s ±  0.666 s    [User: 343.353 s, System: 116.654 s]
  Range (min … max):   79.373 s … 80.671 s    3 runs

Benchmark 2: cd ../scm-prev; dune clean; dune build --cache=disabled
  Time (mean ± σ):     91.555 s ±  0.613 s    [User: 355.560 s, System: 118.064 s]
  Range (min … max):   91.083 s … 92.248 s    3 runs

Summary
  dune clean; dune build --cache=disabled ran
    1.15 ± 0.01 times faster than cd ../scm-prev; dune clean; dune build --cache=disabled

hyperfine --min-runs 2 'dune clean; dune runtest --cache=disabled' 'cd ../scm-prev; dune clean; dune runtest --cache=disabled'
Benchmark 1: dune clean; dune runtest --cache=disabled
  Time (mean ± σ):     103.491 s ±  1.596 s    [User: 374.464 s, System: 125.957 s]
  Range (min … max):   102.363 s … 104.620 s    2 runs

Benchmark 2: cd ../scm-prev; dune clean; dune runtest --cache=disabled
  Time (mean ± σ):     114.158 s ±  2.980 s    [User: 380.638 s, System: 134.558 s]
  Range (min … max):   112.051 s … 116.266 s    2 runs

Summary
  dune clean; dune runtest --cache=disabled ran
    1.10 ± 0.03 times faster than cd ../scm-prev; dune clean; dune runtest --cache=disabled

[edwintorok]

Cherry picked a few more improvements from the old branch, and rearranged dependencies.

dune runtest: 97s, compared to the initial 114s, ~15% improvement
dune build: 67s, compared to the initial 91s, ~25% improvement
koji build time: 276s, compared to the initial 325s, ~15% improvement

[lindig]

If this builds hard to imagine it's not correct.