[TejasNaikk]

Problem

1. Each shard assignment was triggering a separate etcd write, causing excessive watcher triggers on data nodes during node replacement (130+ writes for a single node)
2. Rolling update was updating 4 nodes at once instead of respecting the 20% limit because INITIALIZING shards were incorrectly counted as "converged"

Solution

Batch goal state writes

• Accumulate all shard assignments for a node in memory during orchestration
• Flush to etcd in a single write per node at the end of the cycle
• Reduces etcd writes from O(shards) to O(nodes)

Fix convergence check

• hasActualStateConverged now checks for ShardState.STARTED instead of just shard existence
• Shards in INITIALIZING state are correctly treated as "in transit"
• Uses existing ShardState enum for type safety

Test Plan

Added Unit Tests

Tested in staging

[GeorgeZhai]

LGTM