add CustomFeature base class to support user-defined features with encoding/decoding logic

[alex-hh]

intended as fix for #7220 if this kind of extensibility is something that datasets is willing to support!

from datasets.features.features import CustomFeature

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

[alex-hh]

@lhoestq would you be open to supporting this kind of extensibility?

[lhoestq]

I suggested a fix in https://github.com/huggingface/datasets/issues/7220 that would not necessarily require a parent class for custom features, lmk what you think