This came up while reviewing #59604
and also investigating #59998

There are multiple places in the SQL code where there is an abstraction inversion: the SQL code calls into RPC endpoints in the server package, which are not intended for use directly by SQL.

In particular, these RPC endpoints are not yet defined for SQL-only servers, which occur in multi-tenant deployments.

The proper way forward is to define a separate "sub-service" API inside the SQL package, which depends only on the client-side KV API, and not dependent on the server machinery, and also is guaranteed to be available in SQL-only pods of multi-tenant deployments.

Systems affected:

• SQL list of sessions & queries, and cancellation
• KV-level table size on disk

cc @jordanlewis

Jira issue: CRDB-3176