These were just an artifact of the old unicode concatenation hack and likely
just penalized other kinds of adding. Also, this fixes __(i)add__ on string
subclasses.

 * Optimize unicode_repeat() for a special case with memset()
 * Simplify integer overflow checking; remove the second check because
   PyUnicode_New() already does it and uses a smaller limit (Py_ssize_t vs
   size_t)

 * PyUnicode_Substring() now fails if start or end is invalid
 * PyUnicode_Substring() reuses PyUnicode_Copy() for non-exact strings

 * PyUnicode_Copy() is faster than substring()
 * Fix also a compiler warning

Call PyErr_BadInternalCall() if inputs are not unicode strings.

 * Replace PyUnicode_AS_UNICODE by PyUnicode_AsUnicode, PyUnicode_AS_UNICODE is
   no more a real macro
 * Replace Py_UNICODE by wchar_t in code specific to Windows

 * Use Py_UCS4* buffer instead of Py_UNICODE*
 * Use "I" or "L" format, instead of "u" format

Jarosch.

Jarosch.

 * Change its prototype: PyObject* instead of PyUnicodeoObject*.
 * Remove an old assertion, the result of PyUnicode_READY (_PyUnicode_Ready)
   must be checked instead

Move these macros to unicodeobject.c

Cleanup also the code (avoid the goto).

To make it clearer that Py_BUILD_ASSERT_EXPR(cond) cannot be used as
assert(cond).