Add DeepSpeed to LightningStableDiffusion model

[tchaton]

This PR extracts the DeepSpeed Stable Diffusion Inference Optimization Policy applied over Diffusers models when doing the following.

pipe = diffusers.StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    use_auth_token=hf_auth_key,
    torch_dtype=torch.float16,
    revision="fp16")

print(pipe)
pipe = deepspeed.init_inference(pipe.to("cuda"), dtype=torch.float16)

This PR introduces the following:

• A generic Cuda Graph Replay with dynamic batch sizes and custom ones for Unet, VAE, ClipEncoder.
• A custom policy for Stable Diffusion Unet to replace CrossAttention, TransformerBlock by DeepSpeed optimized one.
• Perform inference in float16
• Add benchmarks scripts to compare HF and our implementation. CF Table: https://www.notion.so/lightningai/Optimising-Stable-Diffusion-5a29e10606884c28b15616108b96236c#5309c54655e34e21a62025db861cf5f1
• Enable to easily switch between 3 schedulers and default to the SOTA: https://github.com/LuChengTHU/dpm-solver

[lantiga]

Looks great, learned a lot :-) I added a couple of comments