This PR adds support for automatic block splitting on read and map transforms, to keep block size bounded to ~500MiB. This avoids potential OOM situations where a map task may consume too much intermediate Python heap memory, or too much object store shared memory for one block.
