Replace with clean non-LoRA model (LoRA VariableState::reset not impl in current NPUW builds)

Synthetic LLM with LoRA adapters (12L, int4, rank=8, GQA 8/2) for NPUW NPU testing

initial commit