feat: improve perf of groupMessagesForPartition (~600x in production use-case)

[benvan]

This PR is the result of diagnosing high cpu usage on our ECS cluster.

It rewrites the src/producer/groupMessagesPerPartition.js implementation to be O(n)
In practice, this results in a significant performance improvement.

In our use-case, we regularly have blocks of many thousands of messages (20k was the scenario that led to this diagnostic).

I'll lead with the goods:

The following runs execute groupMessagesPerPartition with the specified number of messages.

The tests were run on an ECS Fargate instance with 2vCPU allocated.

Running for 10 messages:
old: 0.93ms
new: 0.49ms
improvement: 1.9x

Running for 50 messages:
old: 0.75ms
new: 0.20ms
improvement: 3.7x

Running for 100 messages:
old: 0.63ms
new: 0.21ms
improvement: 2.9x

Running for 500 messages:
old: 1.13ms
new: 0.085ms
improvement: 13.3x

Running for 1000 messages:
old: 4.73ms
new: 0.097ms
improvement: 48.6x

Running for 10000 messages:
old: 135.59ms
new: 3.33ms
improvement: 40.6x

Running for 20000 messages:
old: 1946.62ms
new: 3.23ms
improvement: 602.0x

Running for 50000 messages:
old: 16917.32ms
new: 0.97ms
improvement: 17415.6x