PyPy: Improve negative overflow error messages in int type conversion

reimplement ord() to make its (C long) return value and input argument type (bytes/unicode/char/unichar) available to the type analysis and avoid redundant round-trips into Python space

move declaration of __Pyx_sst_abs() macro to TypeConversion.c instead of more high-level ModuleSetupCode.c

--HG--
extra : amend_source : 69b25b07657205e86aade4f4f8285e721e4648b0

changelog (PyLong optimisations are actually not used in Py2.7 anymore as it has the simpler 'int' type)

tune PyLong optimisations at (gcc) assembly level so that they prefer the most common case of a single-digit

exclude PyLong optimisation for arithmetic operators in Py2.7 as it's actually unlikely that we'd encounter C long sized PyLong objects in Py2.x (which has the 'int' type for that)

fix crash in MarkParallelAssignments transform when it hits a type inferred TypedExprNode that lacks pos/subexprs

extend PyLong optimisations to everything that fits into a C long (Py2.7 uses 15 bit digits even on 64 bit platforms, i.e. C long covers 4 digits)