Split up QuantumOptics by DiffEq dependencies

[amilsted]

OrdinaryDiffEq.jl is now a meta-package that pulls in a bunch of smaller packages, like OrdinaryDiffEqCore, OrdinaryDiffEqLowOrderRK. This splitting was done so that users of OrdinaryDiffEq can pick and choose what they need, avoiding precompilation overhead for unused code.

We could take advantage of this is QuantumOptics.jl, but currently we depend on StochasticDiffEq, which is not split up and pulls in all of OrdinaryDiffEq.

How about we split QuantumOptics.jl and make it a meta-package that pulls in new component packages QuantumOpticsTimeEvolution, QuantumOpticsStochastic, and QuantumOpticsSteadyState?

QuantumOpticsTimeEvolution could have its dependencies trimmed to only pull in the OrdinaryDiffEqCore and OrdinaryDiffEqLowOrderRK components of OrdinaryDiffEq.

This way users who e.g. only want to do Schrödinger and Master evolution can have a much lighter precompile load.

Thoughts? @Krastanov @david-pl @apkille

[Krastanov]

I am all in favor of this, it would be pretty useful for me. Regrettably, I can not volunteer time to perform that change near term myself.

[apkille]

@amilsted I am in favor as well. My bandwidth is a little low for the next couple of months due to other projects. I'd be happy to help contribute to this effort around the end of the year if no one else can.

[Krastanov]

some really good news on this! The upstream sde solvers do not depend on the entirety of OrdinaryDiffEq.jl anymore. Making QuantumOptics.jl compile way faster should be pretty easy without the deep changes and splits suggested here.

See SciML/StochasticDiffEq.jl@7f78540

[Krastanov]

Precompiling StochasticDiffEq@6.72.0 and all of its dependencies

235.236659 seconds (7.17 M allocations: 462.583 MiB, 0.22% gc time, 8 lock conflicts, 1.66% compilation time: 8% of which was recompilation)

Precompiling StochasticDiffEq@6.73.0 and all of its dependencies (which do not include OrdinaryDiffEq anymore)

140.019349 seconds (7.04 M allocations: 449.566 MiB, 0.20% gc time, 8 lock conflicts, 2.84% compilation time: 6% of which was recompilation)

[oameye]

Woop woop

[Krastanov]

Stripping OrdinaryDiffEq leads to 30% decrease in compilation times.

253 dependencies successfully precompiled in 211 seconds. 34 already precompiled.
215.249640 seconds (7.27 M allocations: 469.173 MiB, 0.17% gc time, 14 lock conflicts, 2.43% compilation time: 6% of which was recompilation)

Before the strip:

283 dependencies successfully precompiled in 321 seconds. 34 already precompiled.
325.648877 seconds (7.37 M allocations: 480.778 MiB, 0.10% gc time, 17 lock conflicts, 1.59% compilation time: 6% of which was recompilation)

Next main targets to strip are probably NonlinearSolve and LinearSolve (or whoever is using them)

Here is a PR with a draft of these changes #438