ECQ$^{\text{x}}$: Explainability-Driven Quantization for Low-Bit and Sparse DNNs