This PR introduces a new utility in Accelerator, AcceleratorState, and PartialState: Accelerator.split_between_processes.

It is often useful when performing distributed inference in applications such as stable diffusion to send one prompt to GPU A, another prompt to GPU B, and so forth. This PR introduces a new context manager that let's the user send some data in and split it evenly across all instances for them to use. An example application might look like such:

from accelerate import PartialState
from diffusers import DiffusionPipeline

state = PartialState()
pipe= DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)

with state.split_between_processes(["a dog", "a cat"]) as prompt:
    pipe.to(state.device)
    image = pipe(prompt).images[0]

On a two process system, GPU A would receive "a dog" and GPU B would receive "a cat".

This is also especially useful for cases where using a DataLoader to perform the task is too much code, and the user just wants to send in strings or already preprocessed dictionaries and split them.