On the last line:
# Compute masked accuracy (quotient between the total matches and the total valid values, i.e., the amount of non-masked values)
we do not specify the axis parameter. In other words, we want to calculate all the accurately predicted labels (in our case by summing matches_true_pred which accounts for padded tokens) and divide by the number of all the elements (in our case by summing mask, which accounts for padded tokens). And both of these are scalars (not vectors).
Cheers