Feature request

Would be nice to have a mock worker API without actually using a GPU, so one could do at-scale testings without having at-scale GPUs.

• Need to simulate / predict the KvEvents based on some eviction policy for the worker (feat: vllm mock workers, Rusty skeleton #1033)
• Model the memory + compute complexity based on the active radix cache size and the active request sizes (feat: Add TTFT and ITL Interpolation to Profiling Script #1159), done by @tedzhouhk
• Integrate the above modeling into the mocker. And simulate more closely the vllm scheduling behaviors. Hook it up as an AsyncEngine, or a component that can be registered as a dynamo endpoint. (feat: vllm mocker enhancement #1236 )
• Publishing of RouterEvents and ForwardPassMetrics over NATS, reusing the existing publishers when possible.
• Verify functionality and perform preliminary (sanity check) benchmarking on the KV router.

Describe the problem you're encountering

N/A

Describe alternatives you've tried

A mock worker exists here
https://github.com/ai-dynamo/dynamo/blob/main/components/metrics/src/bin/mock_worker.rs
but seems to be generating random metrics for the purpose of testing metrics handling, and does not seem to perform any modeling