Set default stage in adapter to "inference"

[vpratz]

Concerns #417.
Question:
As Lars already highlighted when creating the current implementation, we cannot pass the stage to the __getitem__ methods of the datasets. How do we want to handle this?
For now, I have made the stage that will be used configurable in the __init__ function.
What should its default be? For now, I have set it to "training". The alternative would be to set it to "inference" and dynamically change it to "training" in Approximator.fit.
What do you think, @LarsKue and @stefanradev93? Setting this to a draft until we have discussed this...

Minor point: Do we want to convert the stages to an enum. This would be a bit more typing (but autocomplete should help), and might prevent future bugs due to typos in the strings. Not sure whether it is worth it though, what do you think?

[codecov]

Codecov Report

Attention: Patch coverage is 75.00000% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
bayesflow/datasets/disk_dataset.py	0.00%	2 Missing ⚠️
bayesflow/datasets/rounds_dataset.py	0.00%	2 Missing ⚠️

Files with missing lines	Coverage Δ
bayesflow/adapters/adapter.py	81.05% <100.00%> (ø)
bayesflow/adapters/transforms/standardize.py	97.43% <100.00%> (ø)
bayesflow/datasets/offline_dataset.py	92.85% <100.00%> (+0.17%)	⬆️
bayesflow/datasets/online_dataset.py	100.00% <100.00%> (ø)
bayesflow/datasets/disk_dataset.py	38.23% <0.00%> (-1.16%)	⬇️
bayesflow/datasets/rounds_dataset.py	36.36% <0.00%> (-1.14%)	⬇️

[LarsKue]

Datasets will pretty much only ever be used directly in training. I am fine with this change, although I do wish we had a better alternative available (such as simply not requiring a stage argument for the Adapter, since it should not be trainable in the first place).

[LarsKue]

I don't think we should change the stage argument to an enum, because either users will then have to make another import (cumbersome), or we make it compatible with strings anyway, in which case this becomes the primary use case.

[stefanradev93]

I suggest we stick to the simplest solution as proposed by Valentin.