fix generation batching and single process accelerator bugs

[adistomar]

This PR fixes two things:

• In generate.py, the state tensor x containing the prompt as well as the masked and unmasked tokens had a hardcoded batch size of 1, meaning that the user cannot perform batching. Fix: set first dim of x to prompt.shape[0] during initialization
• In eval_llada.py, if the # of processes is 1 (for ex. in single GPU case), then self.accelerator is set to None. In this scenario, performing self.accelerator.wait_for_everyone will cause an error. Fix: add a check for if the variable is None