kvserver: limit replica catchup concurrency during Raft scheduling

If many replicas have fallen severely behind on log application, we should consider limiting the number of concurrent replicas doing catchup, to prevent resource exhaustion and scheduling starvation of other ranges. In sum, the overall catchup should take the same time, but this would allow ranges to come online sooner, and prevent e.g. memory blowup across many replicas pulling the full 64 MB command batch into memory at the same time.

Jira issue: CRDB-28229