Currently we parse changesets from download messages all at once into a std::vector of Instructions. This means that we loose a lot of the benefits of the highly space-efficient encoding of changesets during integration which can cause memory pressure on clients. We should make decoding of changesets lazy, so that we decode instructions as needed while iterating changesets, with a cache for interned strings that only stores offsets into the encoded changeset buffer to produce StringData's for InternedStrings.