[REQUEST] add `initialize_inference` wrapper

Is your feature request related to a problem? Please describe.

For Deepspeed ZeRO Inference currently one has to do one of:

ds_engine = deepspeed.initialize(model=model, config_params=ds_config)[0]
ds_engine, _, _, _ = deepspeed.initialize(model=model, config_params=ds_config)

Describe the solution you'd like

I propose a simple wrapper that makes the API slightly more user-friendly for inference users:

# deepspeed/__init__.py
def initialize_inference(*args, **kwargs):
    return initialize(*args, **kwargs)[0]

and then use:

ds_engine = deepspeed.initialize_inference(model=model, config_params=ds_config)

Hmm, thinking more about it, perhaps someone will want the dataloader return value as well, in which case this proposal won't be good.

I have just never used the dataloader side of things in deepspeed, so I'm not sure if someone will want it for inference or not.

If someone may want it then let's close this request.

@jeffra, @tjruwase

Start in deepspeed/init.py and inspect the existing initialize return values and inference usage. Determine whether inference callers need the dataloader return value before settling the wrapper's API. Done means the requested inference entry point has a clear, agreed return behavior.