Evading Data Contamination Detection for Language Models is (too) Easy