Custom imports are broken

[pbontrager]

If you make your own custom dataset.py or model.py and include the import path in the config, these imports will fail because tune run always sets the current working directory to torchtune. Since all of torchtune imports are available anywhere once installed, tune run should use the launch location as the cwd.

Potential complication - since recipes are not in the cwd, we'll have to test how recipe work being launched from a different directory.

[RdoubleA]

Only single device was patched. distributed recipes still do not detect modules from cwd. potential issue: pytorch/pytorch#77737 (comment)