Add supports to iterable-style datasets in adaptdl.torch.AdaptiveDataLoader

[mylibrar]

Currently adaptdl.torch.AdaptiveDataLoader only supports map-style datasets, which allows random access and implements the __getitem__() and __len__() protocols. It might be desirable to add supports to iterable-style datasets since sometimes the dataset might adopt a lazy loading strategy and we cannot know the size of dataset beforehand.