Stop trying to intern arguments in PyObject_FastGetAttrString (pytorch#161432)

If we want them interned, we should intern at callsites. (The numpy reference has bit rotted; see numpy/numpy@b222eb6#diff-6bdb6105198083838f51c57b55b3a49472ed23043bb40018f1ea41138e687163)

Profiling a simple torchdispatch benchmark with perf before/after seems to show that time spent copying std::strings and interning Python strings is gone, though there is some noise and the improvement is very small.
Pull Request resolved: pytorch#161432
Approved by: https://github.com/ezyang
ghstack dependencies: pytorch#161301, pytorch#161292, pytorch#161304, pytorch#161308, pytorch#161315, pytorch#161317, pytorch#161328, pytorch#161329

Author: swolchok

torch/csrc/utils/python_strings.h

@@ -116,7 +116,7 @@ inline py::object PyObject_FastGetAttrString(PyObject* obj, const char* name) {
   }
   /* Attribute referenced by (PyObject *)name */
   else if (tp->tp_getattro != nullptr) {
-    auto w = py::reinterpret_steal<py::object>(THPUtils_internString(name));
+    auto w = py::reinterpret_steal<py::object>(PyUnicode_FromString(name));
     if (w.ptr() == nullptr) {
       return py::object();
     }