run_training not correctly respecting process_data: False

[JamesKunstle]

When setting up the arguments to subprocess torchrun in main_ds.py::run_training, we set the data path to the following:

f"--data_path={train_args.data_output_dir}/data.jsonl",

This assumes that we preprocessed the the input data file and wrote the result to data.jsonl
However, this isn't the case if we've set process_data=False in TrainArgs. If this is the case, we should be honoring the data_path argument as the input .jsonl data file.