Exploring the Impact of Training Data Distribution and Subword Tokenization on Gender Bias in Machine Translation