Adapt vLLM LoRA request path parameter

[amumu96]

Summary

• Adapt vLLM LoRA request construction for vLLM 0.14.0 and newer by passing lora_path.
• Keep lora_local_path for older vLLM versions to preserve compatibility.

Release note

• Fix LoRA module loading with newer vLLM versions that expect LoRARequest(lora_path=...).

Validation

• python -m py_compile xinference/model/llm/vllm/core.py
• git diff --check -- xinference/model/llm/vllm/core.py

[amumu96]

Closed because the PR was recreated against the parent repository: xorbitsai#4980