Address ernie-image review findings #13577

[akshan-main]

What does this PR do?

Partial fix for #13577. Addresses 1, 2, 5 per @yiyixuxu's scope

• (1) Switch ErnieImageAutoPromptEnhancerStep to ConditionalPipelineBlocks so use_pe=False actually skips the prompt enhancer (AutoPipelineBlocks selected on presence, not truthiness).
• (2) Align modular VAE BN epsilon to the standard pipeline's hardcoded 1e-5 (matches training; the hub config currently reports 1e-4).
• (5) Restructure output_type=\"latent\" so it runs maybe_free_model_hooks() and honors return_dict, matching the QwenImage/Flux2 pattern.

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? ernie-image model/pipeline review #13577
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@yiyixuxu

[akshan-main]

https://colab.research.google.com/gist/akshan-main/c6e677d61a865593c1d98aa71aaf2afd/ernie_image_review_fixes_test.ipynb

tested here

[asomoza]

can we also take the opportunity to change the auto classes for the real classes, it gets confusing for some users that pass the text encoder e.g. quantization and also kind of annoying to get the warning all the time.

[akshan-main]

@asomoza switched text_encoder to Mistral3Model and pe to Ministral3ForCausalLM in both the standard and modular pipelines. Left the tokenizers as AutoTokenizer since mistral doesn't have a model-specific tokenizer class

[hlky]

dtype casting to be safe and for consistency with modular

Suggested change

	bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(device)
	bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + 1e-5).to(device)
	bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(device=device, dtype=latents.dtype)
	bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + 1e-5).to(device=device, dtype=latents.dtype)

There could be a TODO regarding vae.config.batch_norm_eps, it should be used in the future if the checkpoint config is changed

[akshan-main]

Added

[hlky]

Can VaeImageProcessor be used here? cc @yiyixuxu Enforcing VaeImageProcessor could be another agent review rule?

[akshan-main]

Switched both standard and modular to VaeImageProcessor.postprocess. Also fixes output_type="pt" in the standard pipeline (was returning numpy).