Remove returned feature constrain

In the current version, the returned value of the map function has to be list or ndarray. However, this makes it unsuitable for many tasks. In NLP, many features are sparse like verb words, noun chunks, if we want to assign different values to different words, which will result in a large sparse matrix if we only score useful words like verb words.

Mostly, when using it on large scale, saving it as a whole takes a lot of disk storage and making it hard to read, the normal method is saving it in sparse form. However, the NumPy does not support sparse, therefore I have to use PyTorch or scipy to transform a matrix into special sparse form, which is not a form that can be transformed into list or ndarry. This violates the feature constraints of the map function.

I do appreciate the convenience of Datasets package, but I do not think the compulsory datatype constrain is necessary, in some cases, we just cannot transform it into a list or ndarray due to some reasons. Any way to fix this? Or what I can do to disable the compulsory datatype constrain?