Implement Executor Level Connection Pool for NodeClient

[ShimonSte]

Currently, each partition task creates its own NodeClient instance, resulting in one ClickHouse connection per partition. When an executor processes multiple partitions (common scenario), this leads to unnecessary connection overhead.

Example:

• 1000 partitions distributed across 10 executors
• Each executor processes ~100 partitions
• Current: 1000 connections created (100 per executor)
• Optimal: 10 connections (1 per executor, shared across partitions)

Current Implementation

// ClickHouseReader.scala
private lazy val nodesClient = NodesClient(part.candidateNodes)
def nodeClient: NodeClient = nodesClient.node