Bandwidth caps for peer replication

[xe-nvdk]

Problem

Catch-up and reactive replication share the worker pool but have no bandwidth limit. A node rejoining with a large backlog can saturate the network.

Proposed Solution

Add a configurable bytes/sec rate limiter on the puller's fetch path. Per-node or cluster-wide.

Context

Deferred from Phase 3. Current mitigation: stagger rejoins, tune `replication_pull_workers` down.

Effort

Medium