Streaming datasets V2

[knighton]

• Sample sizes moved to shards
• V2 small JSON index format (faster startup time at scale)
• Sample fields now have types instead of having to manually provide decoders
• Prefetch factor
• Fast process pool Dataset.download()
• Dataset supports random access of samples, lazily loading their shards
• Offsets live on disk instead of memory for scalability (two seeks instead of one for sample access)
• Various compression algorithms
• Various hash/checksum algorithms on shards (TODO: on index)
• Datasets: CIFAR10, ImageNet, C4
• TODO: plug in the other formats, abstracting Dataset class from MDS

[knighton]

Superseded by #2