Follow-ups for GqlNodeState

Missing functionality from core Rust NodeState:

• Merging. NodeStates can't be merged in GraphQL. Passing another NodeState as argument to the merge function is awkward to express in GraphQL queries.
• Parquet serialization. to_parquet and from_parquet weren't added because they would require filesystem access on the server's machine, which could present a security risk.

Re-implemented:

sum and mean. In Rust, the trait bounds ensure that the inputs to these are numerical/additive. In GraphQL, we can't use these trait bounds to ensure the NodeState has numerical values. As such, these functions were re-implemented and return None if the values aren't numerical/additive at runtime.

Still need to add:

• Paging. Right now, accessing the data in a NodeState is done using columns/nodes/rows. These dump the entire NodeState's contents (columns, nodes, or rows respectively) in a single response, which can be a problem on big NodeStates. Ideally, we can page the results like with other queries.
• Testing for GqlNodeStateValue::Node. It is currently present in GraphQL to match core Rust, but no algorithm returns a single-node column, so it is unused/untested.

Start with the GraphQL NodeState definitions and the existing columns, nodes, and rows accessors. Trace how other queries implement paging, then define completion as paged NodeState results rather than dumping all contents. Also inspect GqlNodeStateValue::Node and add coverage showing its GraphQL behavior.