Add speedups from pyston (#264)

This should make Python go 30% faster. It does that by trading
away some debuggability, like _tracemalloc. It can be re-enabled
using `make MODE=dbg`.

third_party/python/Include/abstract.h

@@ -36,8 +36,12 @@ PyObject *_PyObject_FastCallKeywords(PyObject *func, PyObject **args,
 PyObject *_PyObject_Call_Prepend(PyObject *func, PyObject *obj, PyObject *args,
                                  PyObject *kwargs);
 
+#ifdef USE_CHECKFUNCRESULT
 PyObject *_Py_CheckFunctionResult(PyObject *func, PyObject *result,
                                   const char *where);
+#else
+#define _Py_CheckFunctionResult(func, result, where) (result)
+#endif
 #endif /* Py_LIMITED_API */
 
 PyObject *PyObject_CallObject(PyObject *callable_object, PyObject *args);

third_party/python/Include/cosmo.h

@@ -0,0 +1,9 @@
+#ifndef Py_COSMO_H
+#define Py_COSMO_H
+#include "third_party/python/Include/object.h"
+COSMOPOLITAN_C_START_
+
+PyMODINIT_FUNC PyInit_cosmo(void);
+
+COSMOPOLITAN_C_END_
+#endif /* !Py_IMPORT_H */

third_party/python/Include/pyerrors.h

@@ -318,8 +318,11 @@ void PyErr_BadInternalCall(void);
 void _PyErr_BadInternalCall(const char *filename, int lineno);
 /* Mask the old API with a call to the new API for code compiled under
    Python 2.0: */
+#ifdef USE_BADINTERNALCALL
 #define PyErr_BadInternalCall() _PyErr_BadInternalCall(__FILE__, __LINE__)
-
+#else
+#define PyErr_BadInternalCall()
+#endif
 /* Function to create a new exception */
 PyObject * PyErr_NewException(
     const char *name, PyObject *base, PyObject *dict);

third_party/python/Include/pymem.h

@@ -63,6 +63,11 @@ int _PyTraceMalloc_Untrack(
 PyObject* _PyTraceMalloc_GetTraceback(
     _PyTraceMalloc_domain_t domain,
     uintptr_t ptr);
+#ifndef MODE_DBG
+#define _PyTraceMalloc_Track(domain, ptr, size) (-2)
+#define _PyTraceMalloc_Untrack(domain, ptr) (-2)
+#define _PyTraceMalloc_GetTraceback(domain, ptr) (&_Py_NoneStruct)
+#endif
 
 int _PyMem_IsFreed(void *ptr, size_t size);
 #endif   /* !defined(Py_LIMITED_API) */