Mind the (Language) Gap: Towards Probing Numerical and Cross-Lingual Limits of LVLMs