Fix how DDPOptimizer clones dynamo callback (#107834)
Instead of hardcoding a new callback creation using 'convert_frame',
add an attribute to both callbacks that implement 'self cloning with new
backend', so DDPOptimizer can invoke this in a consistent way.
Fixes #107686
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107834
Approved by: https://github.com/ezyang
