Add experimental support for loading .safetensors models

[monstruosoft]

As the title says, this PR adds experimental support for loading models in safetensors format; I call it experimental because, even if it technically works, I've noticed that rebuilding the pipeline a couple of times, for example by enabling/disabling ControlNet, causes RAM usage to increase considerably making it impractical for its use on low-end PCs.

Some other things to note in this PR:

• Loading safetensors models only works in LCM-LoRA mode, I think it's the only mode where it makes sense.
• LCM-LoRA mode was always considerably slower than plain LCM mode on my machine, about 50% slower, I always thought that it was a speed hit from the LCM-LoRA mode itself but I noticed that by commenting this line: pipeline.unet.to(memory_format=torch.channels_last) inference speed automatically became much better. Is there any particular reason why this line is used?.
• By going further and always fusing the LCM-LoRA with the base model, inference speed in LCM-LoRA mode gets almost as fast as inference speed in plain LCM mode.

[captainerd]

pipe = pipeline_class( TypeError: StableDiffusionPipeline.init() got an unexpected keyword argument 'text_encoder_2'

I tried to load sd_xl_base_1.0.safetensors

It loads fine the same model as diffusers

Am i missing something?

[monstruosoft]

Sorry, should have mentioned that this PR works only for SD 1.5 models. A similar code should work for SDXL but unfortunately I can't run SDXL on my machine.

[rupeshs]

@monstruosoft Going to merge this PR,thanks