Implement faster single batch encoding/decoding for use in shuffle

[andygrove]

What is the problem the feature request solves?

Because Spark shuffle is block-based rather than streaming, it is necessary to serialize single batches along with schema information. We currently use Arrow IPC to do this, but this is not efficient. A crude prototype in PR TBD shows that we can get much better performance with a custom implementation of single batch serde.

Describe the potential solution

No response

Additional context

No response