Improve ratelimit accuracy for stress test situations

[Flo4604]

We need to improve the accuracy of our distributed rate limiting system, as it currently fails under high node and RPS scenarios, exceeding the 20% tolerance threshold when running make test-stress

To resolve this, we should either increase the frequency of how often we read from the Redis counter or use a better method to sync between nodes to ensure they maintain more accurate limits.

Maybe we can use redis streams / pubsub to sync changes between nodes.

[linear[bot]]

ENG-2139