Multi-Action Self-Improvement for Neural Combinatorial Optimization