Simplify API and code by dropping (alleged) support for changing system structure during simulation

[kyllingstad]

In principle, the current libcosim API permits one to change the system structure after a simulation has started. More to the point, one is allowed to call execution::add_simulator(), execution::connect_variables(), etc. after execution::simulate_until() or execution::step() has been called. This was a deliberate design choice in the early OSP days, because we were told that the ability to dynamically add and remove entities was a desirable feature in certain human-in-the-loop simulation use cases.

Now, several years later, I am fairly certain that no one is using this for anything and it doesn't even work. Firstly, because we don't have tests for it, and secondly, for the following reason: If a subsimulator is added with fixed_step_algorithm::add_simulator() after fixed_step_algorithm::initialize() has been called, then the subsimulator’s initialisation functions (simulator::setup() and simulator::do_iteration()) will never be called. This, again, means that the underlying FMU will never have its fmi2EnterInitializationMode() and fmi2ExitInitializationMode() functions called, and then it’s not allowed to call its fmi2DoStep() function either.

This leaves us with two choices: fix this feature or drop it. I say drop it.

Dropping it would allow us to simplify the API in several places:

• An execution could be constructed directly from a system_structure, and we could remove several of its member functions: add_slave(), add_function(), connect_variables(), and set_{real,integer,boolean,string}_initial_value(). Thus, the execution API could be focused solely on actually running the simulation, not setting it up.
• In algorithm, we could remove add_simulator, remove_simulator(), add_function(), connect_variables(), and disconnect_variable(). (Note how the removal functions aren't even exposed via execution yet, and that there is no remove_function(). It's all very half baked.)
• In observer, we could remove the simulator_added(), simulator_removed(), variables_connected(), and variable_disconnected() event functions. None of our observers use the latter two, so all could be replaced by a single setup() function that gets a list of observable pointers.
• For manipulator, the situation is much the same as for observer – no simulator added/removed events needed, just a setup function which gets a list of manipulable pointers.

Furthermore, it would allow us to simplify the implementation in several places:

• The whole inject_system_structure() hack could be removed.
• algorithm, observer, and manipulator implementations would be less complex and potentially more efficient, because their variable mapping tables are set up just once, they don't have to take into consideration that the variable mappings can change dynamically.
• It would make state saving and persistence simpler, because one wouldn't have to keep data related to system structure at various levels (algorithm, execution), only at the outer level or perhaps not at all.

Proposed action plan:

• [ ] Simplify observer and its implementations
• [ ] Simplify manipulator and its implementations
• [ ] Simplify execution
• [ ] Simplify algorithm and its implementations

[davidhjp01]

After discussion, we agreed that we will drop a support on adding/removing simulators (models) and simplify libcosim code. But dynamic adding/removing connections dynamically can be useful (future work).

[kyllingstad]

Pull request #777 dynamically prevents adding/removing subsimulators after initialisation, so it is a step in the right direction. But I think it would be even better to prevent it statically, i.e., by not having functions like add_simulator(), simulator_added() etc. at all. Instead, a complete list of subsimulators could be passed to the execution, algorithm, observers, etc. at some point, for example during construction.