[Stable Diffusion] Fix the image size value inferring

[JingyaHuang]

In the dummy input generator, the default value of image_size would be the sample_size, which would lead to shapes using for compilation unmatch with the height and width given by users.

This PR is suppose to fix how the height and width should be infer.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[JingyaHuang]

Also gently pinging @oOraph and @juliensimon, if you are using optimum-neuron 0.0.9, if the height(width) divide by the vae_scale_factor factor doesn't equal to the sample_size in the checkpoint's config file, the inference might fail due to inconsistent latent shape without the PR.