[C++][Dataset] Separate datasets backed by readers from InMemoryDataset

From ARROW-10882/https://github.com/apache/arrow/pull/9802 

• Backing an InMemoryDataset with a reader is misleading. Let's split that out into a separate class.
• Dataset scanning can then use an I/O thread for the new class. (Note that for Python, we'll need to be careful to release the GIL before any operations so that the I/O thread can acquire the GIL to call into the underlying Python reader/file object.)
• Longer-term, we should interface with Python's async.