We should replicate the existing find benchmarks that use Document to new ones that use the new RawDocument type. We should also introduce benchmarks for borrowed and owned serde deserialization for the find benchmarks. With these benchmarks, we'll be able to measure the performance gains of using the raw BSON types.