[pipelines] fix SD3 crash with pre-computed prompt_embeds and num_images_per_prompt

[zxuhan]

What does this PR do?

Fixes #10712.

StableDiffusion3Pipeline.__call__ raises a RuntimeError when the caller passes pre-computed prompt_embeds together with num_images_per_prompt > 1:

prompt_embeds, neg, pooled, neg_pooled = pipe.encode_prompt(
    prompt="...", do_classifier_free_guidance=True, num_images_per_prompt=2,
)
pipe(prompt_embeds=prompt_embeds, negative_prompt_embeds=neg,
     pooled_prompt_embeds=pooled, negative_pooled_prompt_embeds=neg_pooled,
     num_images_per_prompt=2)
# RuntimeError: size of tensor a (8) must match size of tensor b (4)

Root cause

encode_prompt relies on _get_clip_prompt_embeds and _get_t5_prompt_embeds to apply the num_images_per_prompt expansion. Those helpers only run when prompt_embeds is None, so user-supplied embeddings keep their original batch dimension. __call__ then computes latents = prepare_latents(batch_size * num_images_per_prompt, ...), CFG doubles that to 2 * batch_size * num_images_per_prompt, while encoder_hidden_states stays at 2 * batch_size. The joint-attention block fails when it concatenates the two.

StableDiffusionXLPipeline is not affected because its encode_prompt applies the repeat(1, num_images_per_prompt, 1).view(...) expansion unconditionally after the if/else branch.

Fix

Apply the same expansion to user-supplied prompt_embeds (and the matching pooled_prompt_embeds, negative_prompt_embeds, negative_pooled_prompt_embeds) at the end of encode_prompt. The change is propagated via make fix-copies to the img2img, inpaint, controlnet, and PAG variants:

• pipeline_stable_diffusion_3_img2img.py
• pipeline_stable_diffusion_3_inpaint.py
• controlnet_sd3/pipeline_stable_diffusion_3_controlnet.py
• controlnet_sd3/pipeline_stable_diffusion_3_controlnet_inpainting.py
• pag/pipeline_pag_sd_3.py
• pag/pipeline_pag_sd_3_img2img.py

Adds test_pipeline_accepts_prompt_embeds_with_num_images_per_prompt, which feeds the output of encode_prompt(num_images_per_prompt=k) back into the pipeline. Without the fix the test reproduces the RuntimeError from the issue; with the fix it passes.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? See StableDiffusion3 pipeline RuntimeError when using prompt_embeds #10712.
• Did you make sure to update the documentation with your changes? The existing docstring for num_images_per_prompt already describes the intended behavior; this PR makes the implementation match it.
• Did you write any new necessary tests?

Who can review?

@yiyixuxu @sayakpaul