Improve performance profiling experience #640
Description
When we build the MLIR, we currently encode trace tensor names as the location attributes. We should instead embed the stack information with an option to include code snippets as well so that each kernel in an nsys trace will show us which line of Python code it originates from.