Implementing Data Deduplication Across Simultaneous Queries

As described in ipld/specs#120 it would be great if we could hold off GCing the list of blocks we've sent to a peer until the end of multi-part request. This currently might happen by accident, since as long as one of the queries that accesses a duplicate node hasn't completed yet the reference won't be freed.

Each of the potential approaches mentioned in the issue above has its advantages and disadvantages. However, none of them seem like a particularly huge deal to implement since the protocol is already flexible enough to handle most of it.

@hannahhoward I'm happy to help with the implementation work here as well.