feat: KV-block hashing consistency between vLLM and llmd

[sagiahrac]

Context

This PR updates the LLMD block hashing method to maintain compatibility with a recent change in the upstream vLLM project.

The vLLM block hashing method was updated to use the LoRA adapter name (string) instead of the previous integer ID (vllm-project/vllm#27211). This alignment is necessary for correct KV-cache lookups in LLMD.

Key Changes

1. LoRA Hashing Support: The LLMD block hashing logic is updated to correctly incorporate the LoRA adapter name in the request hash calculation. Previously, LLMD did not support LoRA request hashing.
2. Test Coverage: New tests have been added to verify the behavior of the request hash generation both with and without LoRA adapter information.

Future Work & Notes

• KV Event Hashing: The upstream vLLM project is planning an update to the KV block stored event ([Prefix Cache] Include lora_name in BlockStored event for deterministic KV-cache reconstruction vllm-project/vllm#27577). This will eventually allow us to retrieve all necessary hashing information directly from the events.
• Test Data Source: The KV events used for the current testing were gathered using an offline, local vLLM setup, as referenced in issue [KV-Events] Add vLLM KV cache events demo with ZMQ listener #162.