There is a failing test. It would be good to add a test that fails before, passes after this change.

Is this the same issue in gh-295?

I think it's not the same issue, because gh-295 is about whether to convert from `x` to :obj:`x`.

Now I changed to use a new config variable numpy_role_param. The default value is None, which means "do nothing". The expected behavier can be achieved by setting numpy_role_param = ':emphasis:'. Later I'll add documents and tests.

I noticed that adding a role could break a table structure. In the docstring of numpy.arctan2, with numpy_role_param = ':emphasis:' :

====== ====== ================
:emphasis:`x1`   :emphasis:`x2`   `arctan2(x1,x2)`
====== ====== ================
+/- 0  +0     +/- 0
+/- 0  -0     +/- pi
 > 0   +/-inf +0 / +pi
 < 0   +/-inf -0 / -pi
+/-inf +inf   +/- (pi/4)
+/-inf -inf   +/- (3*pi/4)
====== ====== ================

In fact, I would like to see an example docstrng use case of where this PR is needed.

There are some functions which have size parameter, so `size` appears in some docstrings (For example: polynomial.truncate() ).
These should be interpreted as a parameter (*size* or :emphasis:`size` ), but the existence of the python object numpy.size causes them to be interpreted as the reference.