Port of #1558 Adapt multi_model_api_server to vLLM ServingTokenizatio…

[iboiko-habana]

…n refactor

[Copilot]

Pull request overview

This PR updates the Gaudi v1 integration to track upstream vLLM API changes around serving tokenization/state initialization and KV offload metrics, while also improving HPU KV-cache registration compatibility with upstream NIXL probing.

Changes:

• Update multi-model OpenAI server initialization to use ServingTokenization and the new speech-to-text / pooling factory init entrypoints.
• Extend TensorTuple to expose untyped_storage() and data_ptr() for upstream NIXL KV-cache registration logic, and update LoRA manager creation call signature.
• Refresh KV offloading connector metrics unit tests to match upstream “labeled metrics” payload structure.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
vllm_gaudi/v1/worker/hpu_model_runner.py	Pass vllm_config into LoRA manager creation and add TensorTuple storage/pointer delegation helpers for NIXL KV-cache registration.
vllm_gaudi/entrypoints/openai/multi_model_api_server.py	Port multi-model server state setup to upstream ServingTokenization refactor and new factory-based state initializers.
tests/unit_tests/kv_offload/offloading_connector/test_metrics.py	Update tests for upstream labeled-metrics contract (values nested under label-tuples).