kvclient: DistSender read hedging

If leaseholders struggle to serve a read in time, e.g. due to slow log application, network latency, or node overload, it might be helpful to attempt follower reads on other replicas in parallel, either immediately or after a timeout (configurable).

This would also be useful in combination with follower routing (#93503), which would then eagerly attempt to route around network problems.

Note that this can cause significantly increased load on an already struggling cluster if we send these following a timeout, since we'll be amplifying the read load in a feedback loop.

Jira issue: CRDB-30872