Direct multiple shooting

[PierreMartinon]

Done:

• basic prototype reusing the fixed step schemes from the Collocation discretizer. This basically amounts to having several integration steps with different controls within each time step from the main grid. Uses the updated solution builder in CTModels, with individual time grids for state, control, costate and path constraints. Note that the specific case of 1 internal step is equivalent to the Collocation (same NLP dimensions).

Todo:

• rework control and state parametrization, distinct from ODE scheme. In particular, standardize getters with both 'discrete' (argument: time step index) and 'continuous' (argument: time value) variants.
• variable step ODE solver with proper automatic differentiation (see in particular Compatibility issue with OrdinaryDiffEq due to sparsity analysis in v0.8.3 JuliaSmoothOptimizers/ADNLPModels.jl#280)

[PierreMartinon]

See #561