Deepspeed Launcher

**Describe the feature you'd like**
As LLM fine tuning demand is growing, we see many customers using a very similar framework( Hugginface, Pytorch, Deepspeed) due to added benefits and scale that comes with deepspeed.
**How would this feature be used? Please describe.**
Today customers have to write too much of a boiler plate code to get this working in SageMaker, shh-ing between container, waiting for the setup to make the connection between hosts and be able to then use deepspeed library for training. We belive having this OOTB from sagemaker-training-toolkit will be super useful to speed up the experimentation and model training cycles of scientists

**Describe alternatives you've considered**
Alternative is to build your bespoke scripts and solutions.

**Additional context**
I am an internal customer and currently know atleast 4 different teams ( and growing) working on very similar model.

No files, tests, or entry points are named; start by mapping how sagemaker-training-toolkit currently launches distributed training and reviewing the referenced Hugging Face, PyTorch, and DeepSpeed workflow. Done means an out-of-the-box DeepSpeed launcher removes the SSH, host-connection, and setup boilerplate described here and is usable by the affected training teams.

Written by the indexing model from the issue text.