[ICS] The benchmark result with google_gemma-4-E4B-it-Q4_K_M.gguf model on NPU is not satisfactory

[haarika-madaka]

While testing llama-bench with google_gemma-4-E4B-it-Q4_K_M.gguf model on CPU, GPU and NPU.
The benchmark results on NPU is too low when running tg128 test.

We are able to reproduce the behavior on LNL
./build/ReleaseOV/bin/llama-bench -m ~/models/google_gemma-4-E4B-it-Q4_K_M.gguf