We should expose API that allows users to borrow from the next document in a cursor's result set, enabling borrowed deserialization. This could lead to performance gains when deserializing large documents.

Originally, this work was meant to be done as part of RUST-1039, but it was bumped out due to benchmarks showing the gains weren't significant next to the time spent doing I/O. This has been revisited, and for result sets that are large and have a lot of strings/binary data, borrowed deserialization can make a big difference.