fix(ernie_image): pass attn_mask=None when text is unpadded

[Ace3Z]

Closes #13801.

pipeline.transformer.set_attention_backend("flash") on an ErnieImagePipeline raises ValueError: attn_mask is not yet supported for flash-attn 2., while the equivalent call on ZImagePipeline and Flux2KleinPipeline works fine.

ErnieImageTransformer2DModel.forward was unconditionally building an attention mask (image ones concatenated with the text validity vector) and passing it to dispatch_attention_fn as attn_mask. When every text in the batch is at max length the mask is all ones and isn't doing anything, but the flash-attn 2 backend rejects any non-None attn_mask (the check at attention_dispatch.py:1106).

ZImageTransformer2DModel._build_inputs already does the right thing: it sets attn_mask = None when every sequence is at max_seqlen. Applied the same convention here:

if Tmax > 0 and bool((text_lens < Tmax).any()):
    valid_text = ...
    attention_mask = torch.cat([torch.ones(...), valid_text], dim=1)[:, None, None, :]
else:
    attention_mask = None

ErnieImageSingleStreamAttnProcessor and ErnieImageSharedAdaLNBlock already accept attention_mask: torch.Tensor | None, so this is a transformer level change with no downstream edits.

Added one regression test in the existing test_models_transformer_ernie_image.py. It uses the existing tester config's get_init_dict and get_dummy_inputs (which produce unpadded text by default), intercepts dispatch_attention_fn, and asserts the attn_mask it sees is None. Passes on this branch, fails on main.

For the padded case I ran the forward pass on the same model with the same seed and compared output hashes before vs after the change: byte identical (4442ce01ccddd5bf either way for text_lens=[8,5]). The all-True mask the unfixed code passed is mathematically equivalent to None for attention math, so output is unchanged.

Also stubbed the actual flash-attn 2 backend dispatch check (if attn_mask is not None: raise ValueError) and ran it against the model. On main it reproduces the issue's exact ValueError. On this branch the unpadded path runs clean.