This fixes #3886. Basically, the crash there happens like this:

• When a user calls .get_cpp_params() on an mlpack object in Python, this uses cereal to serialize the C++ object to JSON.
• We then load the JSON object into a Python dict with Python's json.loads().
• We detect any Armadillo objects in the dict and convert them to numpy objects.
• During that conversion, we look for the elem key in the dictionary representation of the Armadillo object; but the code crashes, because for an empty Armadillo matrix (e.g. 0 rows or 0 columns), no elem key is serialized by cereal...

so, I addressed this by checking if the number of rows or columns is 0, and if so, returning an empty numpy object with the correct size.