Adaptive Compression-Aware Split Learning and Inference for Enhanced Network Efficiency