bug: potential rotuer memory leak

Describe the bug

I found that router image memory utilization will steadily go up by running long stability tests. The test is performed using RoundRobin strategy thus no HashTrie and HashTrie nodes are created.

To Reproduce

• Deploy the router image on EKS connecting to a bunch of existing model servers, use the round robin routing strategy
• Enable the EKS container insights thus we can check the pod memory utilization
• Start the long-running benchmark test against the router image endpoint. Currently our test traffic is very predictable and steady. E.g 100 concurrent request. And every user will ask 5 questions. The input context is selected from a fixed pool.
• Check the memory usage of the route pod. You will identify the steady trend of memory usage going up

Expected behavior

The memory usage should go up and down. The usage should be around a baseline per testing.

Additional context

No response

Start by reproducing the long-running benchmark against the router on EKS with RoundRobin routing and monitor pod memory through container insights. Inspect the router's request and routing lifecycle for objects retained across steady traffic, comparing memory behavior with the fixed input pool. Done means memory returns toward a stable baseline instead of steadily increasing.