Networkinference

with sparse arrays, so the input is now converted to a dense array first.
FakeData.linear_in_means() no longer generates an error due to identity being
formatted as csc (but it now generates a warning...).