Antoine, it seem that with encode-leak2.patch, the error path after
PyErr_WarnEx() leaks the value of "v".

I rewrote the whole paragraph to make it more straightforward:
- the normal case is tested first
- all paths end with a "return", and no goto.
- no need to test that PyBytes_FromStringAndSize returns a PyBytes...

I find the code much easier to check in this form, but of course this is
a subjective POV.