Handle nesting for ConvertDType, ToArray, adapt concatenate dispatch

[vpratz]

This PR generalizes the transforms from the Approximator.create_default method, so that they can be used with nested inputs as well. For ConvertDType and ToArray, it supplies a generalization that works on dictionaries.
Edit: For Concatenate, it adds a minor change in the Adapter.concatenate dispatch function to detect the special case of only one key being present also if it is nested in a sequence of length one.

The recursive structure of nested inputs requires some extra utility functions, which I put into bayesflow.utils.tree.

[codecov]

Codecov Report

Attention: Patch coverage is 74.57627% with 15 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
bayesflow/utils/tree.py	63.63%	12 Missing ⚠️
bayesflow/adapters/transforms/to_array.py	85.00%	3 Missing ⚠️

Files with missing lines	Coverage Δ
bayesflow/adapters/adapter.py	85.54% <100.00%> (+0.11%)	⬆️
bayesflow/adapters/transforms/convert_dtype.py	100.00% <100.00%> (ø)
bayesflow/utils/__init__.py	100.00% <100.00%> (ø)
bayesflow/adapters/transforms/to_array.py	80.95% <85.00%> (+1.78%)	⬆️
bayesflow/utils/tree.py	63.63% <63.63%> (ø)

... and 2 files with indirect coverage changes

[LarsKue]

@vpratz Isn't the concat issue already addressed by this if-else in the adapter dispatch?

bayesflow/bayesflow/adapters/adapter.py

Lines 485 to 486 in 449a79a

	if isinstance(keys, str):
	transform = Rename(keys, to_key=into)

[vpratz]

Thanks for the response. I didn't see that, I encountered an error in the transform, probably because I passed a list of length one instead of a string. With that additional context, I think it's better to just add that case in the dispatch as well, and revert the changes in the transform.