Fixes #5837

This is a backward compatibility hack to avoid breaking scipy.sparse
after fixing ravel to respect subtypes. Subtypes are still respected
except in the case of matrices and subclasses of matrices.

This is an ugly hack to preserve backwards compatibility for code
that uses matrices. It is needed since both diag and diagonal have
been changed to preserve subtypes otherwise.

Note that a.diagonal() still returns matrices when a is a matrix.

In PR #5358, np.diagonal was modified to return whatever array type it took in.
Also, np.cumsum and np.clip return the same array type. So, np.ravel's behavior is surprising.
Two tests which were expecting np.ravel to return an array have been changed.
Also, the optional `order` parameter was added to MaskedArray.ravel to make it compatible
(matrix.ravel already had this parameter).

If x is a matrix, np.diag(x) and np.diagonal(x) now return matrices
instead of arrays.  Both of these cause x.diagonal() to be called.
That means they return row vectors (just like x.flatten(), x.ravel(),
x.cumprod(), etc.)

allows passing flags like --pdb to test files
also add call to files where its missing

tostring returns bytes which are not equal to string, so provide a
tobytes function alias.
tostring does not emit a deprecation warning yet so rdepends do not need
to check two names to support older versions of numpy without warnings.