[GPU] Add MoE expert offload-to-disk (OTD) for large MoE models

[zaixing-wang]

Details:

a. Add MOE_OFFLOAD_RATIO GPU plugin property (0–100) controlling the percentage of MoE experts kept resident on device memory. Remaining experts are fetched on-demand from host memory via an LRU cache at inference time.
b. Implement compile-time partial constant allocation: only ratio% of expert weight tensors are allocated on GPU; the rest skip device memory transfer.
c. Add runtime LRU-based expert weight management in the fused MoE 3GEMM+SwiGLU kernel, streaming non-resident experts on cache miss.
d. Enables running large MoE models on GPUs with limited VRAM by trading compute latency for memory savings.

Property Configuration:

// Python (openvino_genai LLMPipeline)
import openvino_genai as ov_genai
pipe = ov_genai.LLMPipeline(model_path, "GPU", MOE_OFFLOAD_RATIO=37)
// keep 37% experts on GPU

// llm_bench (openvino.genai/tools/llm_bench)
// Create config.json: {"MOE_OFFLOAD_RATIO": 37}
python benchmark.py -m <model_path> -d GPU -f ov -t text_gen -ic 128 -lc config.json

Tickets:

Tickets: CVS-184115

AI Assistance:

AI assistance used: yes
GitHub Copilot used for: merge conflict resolution, property rename refactoring (MOE_OFFLOAD_MAX_EXPERTS → MOE_OFFLOAD_RATIO), code refactoring of the MoE OTD runtime path, evaluating and selecting file I/O strategies for weight streaming (mmap vs ReadFile vs buffered read), PR branch preparation (squash commit), and drafting this PR description. All changes validated by full rebuild + end-to-end benchmark.