• Consider making the fill value an additional parameter (potentially with the default value of None), to let callers decide what value to use to "normalize" the rows. None might be not always a suitable value and in some cases, some callers would want to use something else.
• Decide how to handle mutation of input, and enforce it properly with tests. I would strongly recommend forbidding mutation of the original array and properly testing against it, because it makes no sense for a transposition operation to modify (and destroy) its input. Additionally, mutated input makes assertion messages confusing.

You should specifies the range, it is greater than 0 or less than 0, it is important.