Describe the bug

It is possible to register custom features using datasets.features.features.register_feature (#6727)

However such features are not compatible with Features.encode_example/decode_example if they require special encoding / decoding logic because encode_nested_example / decode_nested_example checks whether the feature is in a fixed list of encodable types:

This prevents the extensibility of features to complex cases

Steps to reproduce the bug

class ListOfStrs:
    def encode_example(self, value):
        if isinstance(value, str):
            return [str]
        else:
            return value
feats = Features(strlist=ListOfStrs())
assert feats.encode_example({"strlist": "a"})["strlist"] = feats["strlist"].encode_example("a")}

Expected behavior

Registered feature types should be encoded based on some property of the feature (e.g. requires_encoding)?

Environment info

3.0.2