I wonder if a patch for ctypes like this (which is not yet complete)
could be used to implement this, or MUST it be implemented in C?

The patch contains a mixin class that implements the numeric methods. 
However, the actual operation takes place in Python; only the operands
are converted into ctypes types first, the operand is applied to the
value, and the result is converted to a ctypes instance again.

One difficulty with the patch is that the original ctypes code contained
a tp_as_number member where only the nb_nonzero slot was actually
implemented; this prevented the mixin class to do it's work (Python
didn't call the special methods.  I wonder if there is a bug somewhere...).