Korean neural sparse encoder: SPLADE-max with ModernBERT backbone

Update to V26: Recall@1 40.7% (+44% vs V25)

v22.0: InfoNCE contrastive loss, temperature annealing, 99.87% recall