Fix GenAI server LLM inference on SOC_ACCELERATOR devices (Hailo-10H)

[jordanskole]

Summary

• Fix HAILO_NOT_IMPLEMENTED during model creation on SOC_ACCELERATOR devices by using the MemoryView overload of create_infer_model(), which correctly routes to
  InferModelHrpcClient on VDeviceHrpcClient
• Increase WAIT_FOR_OPERATION_TIMEOUT from 10s to 120s to accommodate large HEF transfers over PCIe RPC (e.g., ~2.3 GB for Qwen2.5-1.5B on Raspberry Pi 5)
• Fix use-after-free in TokenEmbedder where the Eigen::Map backing memory (from the HEF buffer) is freed after model creation completes, causing SIGSEGV during token
  generation. Call set_resource_guard(hef_buffer) to keep the buffer alive. This bug affects all device types, not just Hailo-10H.

Files changed

• genai/llm/llm_inference_manager.hpp / .cpp — pass hef_buffer through, use MemoryView overload
• genai/llm/llm_server.hpp / .cpp — thread hef_buffer through creation chain, call set_resource_guard()
• genai/vlm/vlm_server.hpp / .cpp — same fixes for VLM path
• genai/utils.hpp — increase WAIT_FOR_OPERATION_TIMEOUT

Test plan

• Verified LLM inference works end-to-end on Hailo-10H (RPi 5) with Qwen2.5-1.5B-Instruct
• Model loads without HAILO_NOT_IMPLEMENTED or HAILO_TIMEOUT
• 1000+ tokens generated without SIGSEGV