As discussed back at promcon, we should add a map[string]string with hints like the function wrapping the selector, the range and the step. This is unstructured and experimental as we don't know yet what if any of this will prove useful for things like downsampling optimisations on the other end.

At some point (likely after the rest of remote read is considered non-experimental), this should be changed to proper structured data when we have a good idea from multiple implementations what information is needed.