fix(ernie-image): pass attn_mask=None when text is unpadded

[Ace3Z]

What does this PR do?

Fixes #13801.

If you call pipeline.transformer.set_attention_backend("flash") on an ErnieImagePipeline, it crashes:

ValueError: attn_mask is not supported for flash-attn 2.

The same code works fine on ZImagePipeline and Flux2KleinPipeline.

Root cause

ErnieImageTransformer2DModel.forward always builds a boolean attention mask from text_lens, even when every sample already has the full text length. In the common single-prompt case (where Tmax = lens.max() in _pad_text), every position is valid and the mask is just all True. flash-attn 2 rejects any non-None attn_mask, so the call fails before any attention runs.

Z-Image already handles this. In _prepare_for_attention (transformer_z_image.py:792-797) it does:

if all(seq == max_seqlen for seq in item_seqlens):
    attn_mask = None
else:
    attn_mask = ...  # build it

Fix

Do the same thing in Ernie's forward. If at least one sample is padded, build the mask. Otherwise pass None.

The previous all-True bool mask was a no-op on sdpa, cudnn and native paths (no positions get masked, no -inf rows), so behavior on those backends doesn't change. I verified that numerically: with the same seed and inputs, baseline and fix produce bit-identical output in both the uniform (text_lens=[16,16]) and padded (text_lens=[16,12]) cases. Norm, mean, std, and per-element samples match to 7 decimal places.

flash-attn and its variants are now usable when the batch is uniform.

Test

Wires the existing AttentionBackendTesterMixin into the Ernie test class (mirrors the Flux pattern). Native_cudnn backend tests pass; flash variants skip cleanly when kernels isn't installed, and will catch the regression when it is.

Before submitting

• This PR fixes a typo or improves the docs.
• Read the contributor guideline.
• Read the philosophy doc.
• Discussed via GitHub issue (Incompatibility between FlashAttention and ERNIE Image #13801).
• Documentation: mask construction is internal to the forward and not referenced in docs/source/en/.
• New tests: wired in AttentionBackendTesterMixin for Ernie.

Who can review?

@yiyixuxu @sayakpaul

[Ace3Z]

Friendly ping on this one. @yiyixuxu @sayakpaul, when you have a moment, would you mind taking a look? It's a small 10 line change to transformer_ernie_image.py plus a 5 line test wiring in the existing AttentionBackendTesterMixin, mirroring the Z Image pattern. Closes #13801.