ArithmAttack: Evaluating Robustness of LLMs to Noisy Context in Math Problem Solving