make FEMap a gufe tokenizable

[richardjgowers]

To allow serialisation between sessions

[richardjgowers]

I started playing with this, and one big change is that we'd have to make the FEMap object immutable. Is this an ok choice for this? @dwhswenson opinions?

The alternative would be we'd just make the to_dict/from_dict not via gufe.

[ijpulidos]

Do we need to make it gufe tokenizable or just serializable? I think we would like to make the object easy to extend for the "live networks" plan that we have for alchemiscale and getting results from it. I guess this could be done with immutable objects (creating a new one every time), but may not be necessary.

[dwhswenson]

I'm not sure enough about the use cases here. GufeTokenizable adds a few potential conveniences (versioning of serialization, ease of use when saving to gufe-based storage, a few useful points on provenance, etc). But I'm not familiar enough with the needs of cinnabar users to say whether that is worth dropping mutability.

I'll argue strongly for immutable objects if they're likely to be stored as part of a larger database. If the serialization is always directly to a single file for that object, I feel like mutability causes fewer problems (it is understood that you're overwriting the old file). However, when you save into a larger database, mutability requires that the database manage updates, and you lose provenance tracking. That mainly matters if you're using objects that are then themselves used as inputs elsewhere; I'm not sure how much that would happen with an FEMap.

[richardjgowers]

Immutable is a bit misleading, it just means that the pattern for updating a network behaves differently

# mutable
m = FEMap()

for line in myfile:
    m.add_absolute_measurement(line)

# immutable
m = FEMap()
for line in myfile:
    m = m.add_absolute_measurement(line)

Maybe a simple to/from dict is enough for FEMap though