Rethinking Pruning Large Language Models: Benefits and Pitfalls of Reconstruction Error Minimization