There should be a single connection to the "outside"

The pipeline should be clearer.

We should have a single "final" function, i.e., the last function in the pipeline, that will call `solve!`. This `solve!` call should resolve to the specific solver definition, while all others should dispatch to internal calls.
Right now there are at least three places that call external functions:

- `minimize` in `solve.jl`
- `solve!` in `solve-model.jl`
- `minimize` in `ipopt_solve.jl`

To simplify argument passing, we should focus on a simpler pipeline. Something like

1. minimize(args...; kwargs...) # Maybe some variants
2. minimize(solver, nlp; kwargs...) with no variants
3. solve!(solver, nlp, ...) external call

Then internally, we can a new function `_create_problem` that dispatches on the arguments, instead of `minimize`.

Then we have function 1 to create `nlp`, call `select_optimizers`, define the solver, and call function 2.
Function 2 serves mainly to define our interface to call `solve!` (function 3).

Maybe we'll need an intermediary function.