Use vLLM Native Simulator for KV-Cache simulation

[mayabar]

The vLLM Native Simulator is currently available only as an unmerged PR

At the moment, inference-sim supports two modes:

• No KV‑cache simulation
• Internal KV‑cache simulation using an LRU eviction strategy

Add a new option to allow using the vLLM Native Simulator as the KV‑cache implementation.

Ensure that this new mode correctly supports:

• Data parallelism (DP)
• KV‑events replay