[NPUW] Fix MoE model builder for NPU compatibility
- Router weights use i8 (not user-specified i4/nf4), matching real
GPT-OSS two-pass quantization where router is excluded from 4-bit
- Use i32 for MoE shape/axis/step constants matching HuggingFace export
- 3D expert scales [batch, rows, 1] for DEVICE_ROUTED expert detection
- Router nodes named .mlp.expert.router.* for DEVICE_ROUTED transform
Related: EISW-209517
