[BE] Unify `buffer_size` across datapipes

The buffer_size parameter is currently fairly inconsistent across datapipes:

Here are my suggestion on how to unify this:

• Use the same default buffer_size everywhere. It makes little difference whether we use 1e3 or 1e4 given that it is tightly coupled with the data we know nothing about. Given today's hardware / datasets, I would go with 1e4, but no strong opinion.
• Give every datapipe with buffer the ability for an infinite buffer. Otherwise users will just be annoyed and use a workaround. For example, torchvision simply uses INFINITE_BUFFER_SIZE = 1_000_000_000, which for all intents and purposes lives up to its name. Which sentinel we use, i.e. -1 or None, again makes little difference. I personally would use None to have a clear separation, but again no strong opinion other than being consistent.
• Do not warn on infinite buffer sizes. Especially since infinite buffer is not the default behavior, the user is expected to know what they are doing when setting buffer_size=None. I'm all for having a warning like this in the documentation, but I'm strongly against a runtime warning. For example, torchvision datasets need to use an infinite buffer everywhere. Thus, by using the infinite buffer sentinel, users would always get runtime warnings although neither them nor we did anything wrong.

Compare the buffer_size implementations for Demultiplexer, Forker, Grouper, Shuffler, MaxTokenBucketizer, UnZipper, and IterKeyZipper. Review the linked torchvision infinite-buffer example, then determine whether the project can agree on one default, one infinite-buffer sentinel, and no runtime warning; done requires consistent behavior across all listed datapipes.