Refactor bonded interactions storage

[biermanncarl]

Fixes #4225
Fixes #4377
Fixes #4169

Description of changes:

• Bonded interactions are now stored as ScriptInterface objects, are immutable and can be removed
• It is no longer possible to overwrite a bond object by a different type of bond (avoids segmentation faults)

[jngrad]

Thanks a lot! From my side, this is ready for merge.

[biermanncarl]

Also fixes #4169 ?

[RudolfWeeber]

Thank you, Carl. This is a big step forward.

I think we can merge this now, also to unblock related work.

However, some improvements should still be considered:

• It looks like the BondedInteraction script interface class still knows about the numerical id (i,e, key) of the bond in the BondedInteractions storage. That should be avoided. The individual bonded interaction should not (need to) know, how it is stored.
• At least the valid_keys() Python function can probably be autogenerated from the parameter list of the script object.

[jngrad]

At least the valid_keys() Python function can probably be autogenerated from the parameter list of the script object.

Good point! The valid_keys() and required_keys() methods are now deduced from the script object.

It looks like the BondedInteraction script interface class still knows about the numerical id (i,e, key) of the bond in the BondedInteractions storage. That should be avoided. The individual bonded interaction should not (need to) know, how it is stored.

This has to remain so for now, unfortunately. The BondedInteractions container serializes the bond objects by storing the hashmap key (aka bond id) and the bond object parameter dict in the pickle file. During deserialization, a new bond is created from the parameter dict and it is stored in the hashmap with the same key. So in principle we don't need to store the key in the bond object itself, however the bond objects are also stored as a copy in the ParticleHandle objects, and the only way to make the particle bond lists agree with the system bond list after deserialization is to store a unique identifier in the bond objects, e.g. the bond id.

This is not a good design, because each bond is serialized and deserialized multiple times. There is also the issue of the IBM bonds that are hiding the particles ids, and therefore cannot be checkpointed (this is already an issue in the current python branch). I tried to work around that by re-implementing the script interface hashmap in python (4f509f1), but this adds more complexity and doesn't work anyway, because IBM bonds as a side effect do a particle fetch in the constructor, but there is no guarantee that all particles in an IBM bond have already been loaded by the time the first particle in the bond is being loaded.

We can probably revisit this design later, for example in the planned project to store bonds symmetrically on particles.