Delay importing torch & friends

time python -c 'import cuda.tile'

Before/after: real 0m1.142s / 0m0.054s

Additionally, optimize the steam argument parsing in launch(). Adding a
simple cache shaves off about 50 nanos.

Signed-off-by: Greg Bonik <gbonik@nvidia.com>

Author: gbonik

cext/py.h

@@ -244,6 +244,10 @@ void log_python_error(const char* filename, int line, const char* level, SavedEx
 #define LOG_PYTHON_ERROR(level, exc, ...) \
         log_python_error(__FILE__, __LINE__, level, exc, __VA_ARGS__)
 
+static inline PyPtr getattr(PyObject* obj, PyObject* attrname) {
+    return steal(PyObject_GetAttr(obj, attrname));
+}
+
 static inline PyPtr getattr(PyObject* obj, const char* attrname) {
     return steal(PyObject_GetAttrString(obj, attrname));
 }