Exploring Internal Numeracy in Language Models: A Case Study on ALBERT