Deprecate old serialization ((de)serialize_value_or_type) and add developer docs

[vpratz]

The functionality was replaced by the more powerful serialize and deserialize functions. Therefore I would suggest we deprecate the old functions.

It was still use in the point estimation setting, this change would break loading of previously saved models that use those classes. I just noticed that the serializable adapter is currently producing wrong packages (I will open and link an issue soon). I first thought that it was related to my changes, and already fixed it in a1b4d19. But the problem is also present on dev and main.
As the fix for this also breaks backwards compatibility, I think doing this deprecation now might be a good.

This PR is a continuation of #449 which also contains the fix, so the diff will simplify once #449 is merged (or we merge both jointly by merging this one).

[codecov]

Codecov Report

Attention: Patch coverage is 86.95652% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
bayesflow/utils/serialization.py	75.00%	3 Missing ⚠️

Files with missing lines	Coverage Δ
bayesflow/networks/point_inference_network.py	78.94% <100.00%> (-0.44%)	⬇️
bayesflow/scores/scoring_rule.py	93.61% <100.00%> (ø)
bayesflow/utils/__init__.py	100.00% <100.00%> (ø)
bayesflow/utils/decorators.py	92.30% <100.00%> (+1.39%)	⬆️
bayesflow/utils/serialization.py	90.69% <75.00%> (+0.13%)	⬆️

[vpratz]

I cannot get the PointInferenceNetwork to work with the new functions. @han-ol Do you have time to take a look at this?
Edit: Not making other changes to the serialization in the classes helped, I think this is ready to be reviewed.

[han-ol]

Thanks for looking into this and glad that it worked out. With respect to serialization changes of PointInferenceNetwork I have nothing to complain.

I take it you also tried to address the awkwardness documented in https://github.com/bayesflow-org/bayesflow/blob/main/bayesflow/networks/point_inference_network.py#L80. Do you think we can make progress here before the upstream keras issue is resolved?

In any case, this PR is ready to merge after @LarsKue takes a look.

[vpratz]

Thanks for your comments. Regarding your question: Yes, in my first attempt I tried to streamline it a bit, but it failed with error messages I could not learn a lot from, unfortunately. They might well be related to the issue, but I did not explore it further. As the new functions do a very similar thing as the old ones, my guess would be that we cannot resolve this yet.

[LarsKue]

Thank you for the PR and for the added docs!