Add optimization_type property for structures

[rartino]

This is a proposed solution for #406

Before you fire up your keyboards to comment on all the many ways this can/should be extended to allow other possible aspects representing the full provenance of experimental, computational, and AI-model-created structural information, please realize that this property isn't exactly designed to do that.

This is meant to be "a flag" to address a very real and quite pressing need - which we have experienced ourselves when creating OPTIMADE use examples - to have some relatively simple way to filter out "random noise" when querying the OPTIMADE network for "all structures". Of course, what is random noise for one person is signal for another. I've tried my very best to walk the delicate balance to define a few reasonable categories that, based on the discussions in #406, covers actual use cases.

I'd be more than happy to see (and review) a future PR for representing a much more rich account of all the experimental and computational history that led to arriving at some structural information. But, I also believe that representation can co-exist with the "flag" proposed in this PR.

[ml-evs]

Thanks for writing this up @rartino.

I have quite a few comments below but definitely still support the overall idea.

[rartino]

@vaitkus @ml-evs Thanks for reviewing this. I've now tried my best to accommodate all comments, except the rename "processed" -> "derived" which I think we need to discuss some more.

[ml-evs]

I'm happy with all the wording here now, modulo processed vs derived etc (but happy to bow to consensus). I guess this is one of our only recent scientific additions to the data model, so we should probably present it at the next meeting for feedback before merging? If we are happy with the current state of this PR (I think we should be) then perhaps we could even advertise it and ask each database to take a look before the meeting.

[rartino]

It makes sense to present it on the web meet. I'd be happy to yield to the greater consensus between "processed", "derived", or any other word the swarm intelligence can come up with.

[rartino]

@sauliusg @ml-evs @vaitkus @JPBergsma For the changes we discussed at the web meet, I think the "non-extreme conditions" actually only apply to the category predicted, so I moved it there. I think this works; please review.

The whole text is rewritten

[rartino]

Ok, lets make another attempt at this.

I've now completely reworked all text of this PR to adhere to the alternative framing of classification suggested by @sauliusg here: #455 (comment)

I've demoted the PR to a draft, because it makes sense to first iterate this attempt with @sauliusg to make sure it fulfills the goals described in his comment before it is moved to others' review.

[sauliusg]

We could add an example of a specific minimised function(s) (e.g. \sum_{hkl} (F^{obj}{hkl} - F^calc}{hkl})^2 -> min for experimental single crystal structures); but this is probably too specific, so for now I would leave the current wording with the example of @rartino.

[rartino]

On the web meet it was decided to slightly adjust the formulation of local to clarify that it is ok to give that label if a structure is locally optimized but may be globally optimized, it just isn't know.

[ml-evs]

Meant to do this before the meeting yesterday, but I am happy with this classification and think the text is clear.

Ideally we should merge this and try to get some databases to apply the categorization before the next release.