[Flux with CFG] add flux pipeline with cfg support

[linoytsaban]

In the recent PuLID Flux adaptation, authors (ToTheBeginning Yanze Wu, guozinan126, guozinan126) also introduce true CFG - a simple & straight forward way to support negative prompts and "real" CFG for the distilled flux models.

Based on their idea and code, this PR adds a flux pipeline supporting negative prompts.
Inference can be done as follows -

import torch
import diffusers
from pipeline_flux_with_cfg import FluxCFGPipeline

pipe = FluxCFGPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", 
                                    torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()

prompt = "a watercolor painting of a unicorn"
negative_prompt = "pink"


#### without true cfg
torch.manual_seed(0)
img_1 = pipe(prompt).images[0].resize((512,512))


#### with true cfg
torch.manual_seed(0)
img_2 = pipe(prompt=prompt, negative_prompt=negative_prompt, true_cfg=1.5, guidance_scale=3.5).images[0].resize((512,512))

output examples courtesy of @apolinario 🤗:

[yiyixuxu]

thanks!

[arnold408]

Hi,
Will this work with Flux.1-schnell ? What should be the "guidance_scale" ?
Thanks