Major Refactoring into API and Subpackages

[AlCap23]

Supersedes #355 and probably #363 .

• [x] Build the basic functionality à la SciMLBase

Add subpackages for

• [x] Sparse Regression
• [x] Koopman
• [x] SymbolicRegression
• [x] Adapt the documentation accordingly

~~ OccamNet - This will be an individual PR given that I adapt for Lux.jl~~

[AlCap23]

The base functionality should be relatively stable now ( give or take ).

• Using the InternalDataDrivenProblem all preprocessing steps are handled within CommonSolve.init.

• All problems are lifted with a basis, within SymbolicRegression this can be used to precompute features.

• Normalisation and Data Split is sourced out to StatsBase and MLUtils.

• Preprocessing of the fit targets can be handled before the split via get_fit_targets, which allows to dispatch for Koopman Generators

• The structural information about control and implicit variables is stored in the InternalDataDrivenProblem and can be used from Koopman and ImplicitSparseRegression algorithms

• The Basis now has a dedicated Struct for their generated functions , DataDrivenFunction, which allows for better behaviour and no anonymous functions

• The Basis know about controls and implicits -> dispatch pattern

• The Dataset bug has been fixed

• Parameters and variables get now created in accordance to MTK

• The DataDrivenResult extends StatsBase functionality for AIC, BIC etc.

The main API solve(problem, [basis], alg; options) is relatively untouched. I need to convert possible kwargs to DataDrivenCommonOptions, but this is it.

These are hard, but necessary changes and probably a step in the right direction ( hence v1.0 from here on )

Note All invalidations come from Deps here. I might descent into madness to investigate this further

[AlCap23]

Strange note:

When running the pendulum tests locally, I get the expected result from Pkg.test() iff I include the info print. If not, the test fail. This seems super strange. If I include it, everything works out. Maybe I should wrap everything inside a safetestset. Lets see what the CI holds here.

Edit: I relaxed it a little. Might be the RNG.

Additionally, Implicit now works reliable and rejects results without implicit variable dependencies. This discards results like 1 = sin^2 + cos^2.

[AlCap23]

LGTM. Will merge in approx. 8 hrs.