Optimize document store and sync algorithm

The sync algorithm currently reads large parts of the store in to memory. E.g.

// TODO: This is hugely inefficient and needs to be optimized
// For an identity range that includes everything we allocate a vec with all entries of
// the replica here.
let local_values: Vec<_> = self
    .store
    .get_range(range.clone())?
    .collect::<Result<_, _>>()?;

this will be fine for a few 1000 entries, but it won't scale for really large documents. We should make sure that iroh provides few but load-bearing primitives, and currently iroh docs (unlike e.g. blobs) are not load-bearing. As soon as you build something big on top, they will crumble.

I think we can come up with a trait for the store that allows implementing an efficient mem store and disk store, and refactor the sync algo (process_message) to use this trait.

The trait should

• make it possible for the store to cache fingerprints (the current one does this as well)
• provide an interface that allows pivot selection without loading all keys and values into mem
• be sync (no async here, since that will be hell. In the worst case we drive this thing from an actor that has it's own thread)
• support efficient prefix replacement like in the original willow design