feat: implement per-partition locking for parallel throughput

[fuziontech]

Summary

Replace the single global ReentrantLock with per-partition locks to allow parallel processing across partitions.

Problem

The connector had a single global lock that serialized ALL operations:

• With 16 tasks, only 1 could process at a time
• This caused severe lock contention warnings: "Time-based flush check skipped 200+ consecutive times"
• Throughput was bottlenecked by the single lock
• Consumer lag was growing instead of shrinking (~18.6 billion messages behind)

Solution

Implement per-partition locking:

// BEFORE:
private final ReentrantLock flushLock = new ReentrantLock();

// AFTER:
private final ConcurrentHashMap<TopicPartition, ReentrantLock> partitionLocks = new ConcurrentHashMap<>();

Changes

• Replace single flushLock with ConcurrentHashMap<TopicPartition, ReentrantLock>
• Update put() to group records by partition first, then acquire per-partition locks
• Update checkTimeBasedFlush() to iterate partitions with per-partition locks
• Add per-partition buffer methods
• Track flush skips per-partition instead of globally
• Update stop() to acquire per-partition locks during shutdown

Expected Impact

• Before: All 16 tasks serialize on one lock
• After: Each partition has its own lock - parallel processing
• Expected improvement: 2-4x throughput

Test Plan

• Code compiles
• Existing tests pass
• Deploy and verify reduced lock contention
• Verify consumer lag decreases

Fixes #2

🤖 Generated with Claude Code