[intel-npu] refine microsoft matmulnbits op, to fit NPU mix precision pattern

[bopeng1234]

Details:

• when MatMulNBits met requirements of NPU (sym INT4, channel-wised Quantization) the performance raised a lot

    // For NPU Optimization
    // reference: https://github.com/openvinotoolkit/openvino.genai/tree/master/samples/python/text_generation#npu-support
    //  1. model must be exported with symmetric INT4 quantization.
    //  2. The quantized LLM MatMul Op need to use Channel-wised quantization for better performance. (block_size = K)
    // with the limitation, the NPU vpux-plugin compiler could use mix-precision matmul with i4 weight input.
    // it raised the performance a lot in NPU.
  

This PR is going to optimize the decomposition OPs (convert/multiply/matmul) to match the NPU optimize pattern, before this PR, the INT4 phi3 onnx model using NPUW runs:
1st token 8000ms, 2nd token 9040ms, 3rd+ token avg 4000ms

apply this PR, the same model could reach to:
1st token 3400ms, 2nd token 3400ms, 3rd+ token avg 250ms

16x speed up in NPUW for phi3 model