A computation graph for OCPs.

[ManifoldFR]

A problem encountered in #133, which was supposed to solve #79, was how to properly define a computation graph for optimal control problems: how do operations (matrices, vectors, "operators") flow between costs, dynamics, constraints and the functions and objects they call into from other libraries (e.g. Pinocchio).

Having such a graph would enable:

• easily creating some kind of autodiff
• shared computation Sharing data between costs, constraints, dynamics #79

Discussing with others, we have come at the conclusion a computation graph would look like some of DAG with some extra structure/constraints (due to some nodes necessarily being arranged as a timeline, others necessarily being cost functions and dynamics).

This would address the following issues:

• Sharing data between costs, constraints, dynamics #79
• Parametric problems and differentiability #80
• Validate function size before running the solver #117