Description of feature

We currently have three issues (#637, #649, #662) dealing with how we determine feature types and the downstream problems that arise due to different approaches. I'll provide a summary of these issues and associated ToDos here, while closing the other three issues.

Problem:
We've introduced PR #697 to accurately determine feature types in ehrapy. With the new method ep.ad.infer_feature_types, feature types are guessed based on predefined rules and we prompt the user to review these annotations. Currently, feature determination occurs at multiple stages inconsistently and is saved in the adata at several places. Ideally, we would harmonize ehrapy to use exclusively use ep.ad.infer_feature_types for feature annotation, eliminating guesswork in downstream analyses. This means that the new method would be part of the standard preprocessing steps.

ToDos:

• Update encoding: The autodetect option of encoding here relies on the adata.var[EHRAPY_TYPE_KEY] tag, which is set when (1) reading a dataframe here or (2) moving something from obs to X here. Ideally, we would get rid of the adata.var[EHRAPY_TYPE_KEY] annotation entirely and, if autodetect is set to True in the encoding method, base the identification of features to encode on the annotation from ep.ad.infer_feature_types.
• Ensure that the feature type annotations are not stored elsewhere. For example, during encoding, we store the information in adata.uns["var_to_encoding"] and adata.uns["encoding_to_var"].
• Update CohortTracker to be based on ep.ad.infer_feature_types.
• Update ep.tl.rank_features_groups to be based on ep.ad.infer_feature_types.
• Update all methods in adata_ext to be based on ep.ad.infer_feature_types.
• Ensure functionality for both encoded and unencoded data, so it can be used when loading an encoded dataset (I don't know why this wouldn't be the case, still I want to double-check).
• Improve date detection in ep.ad.infer_feature_types to also work with date(time)s stored as strings and update the FHIR tutorial accordingly.
• Decide how to deal with ordinal categorical features: Should we have explicit annotations for ordinal categorical data? In downstream analyses, these could either be treated as a continuous integer-scaled feature, or as a nominal class feature, depending on the context and analysis. (see Distinguish between ordinal and nominal categorical feature types #713)
• Look into whether we can combine this with the feature type detection of TableOne here
• Discuss whether we want to automatically incorporate the feature type detection into the dataset-loaders (i.e. without the user specifically calling it, which would be required to maintain the encode parameter for the data loaders)
• Take care of df_to_anndata, which also does a lot of type inference.
• If we don't fully automate the feature type detection, it might be good to add that to the dataloaders that we maintain