fix: pass attention_mask to T5/CLIP encoder in HFEmbedder

[Mr-Neutr0n]

Summary

HFEmbedder.forward() tokenizes text with padding="max_length", which pads short sequences to max_length with padding tokens. However, it then passes attention_mask=None to the underlying T5/CLIP model.

For T5 in particular, this is problematic because T5 uses bidirectional (fully visible) self-attention. When attention_mask=None, the encoder treats every token — including padding tokens — as real input and attends to them. This means padding tokens actively participate in self-attention computations, polluting the text embeddings especially for short prompts that have many padding tokens.

The tokenizer already produces the correct attention_mask in batch_encoding, marking real tokens as 1 and padding tokens as 0. This PR simply passes that mask through to the model instead of discarding it.

Changes

• src/flux/modules/conditioner.py: Replace attention_mask=None with attention_mask=batch_encoding["attention_mask"].to(self.hf_module.device) in HFEmbedder.forward()

Impact

• Short prompts will produce cleaner text embeddings since padding tokens will be properly masked out
• Long prompts that fill the full max_length are unaffected (no padding tokens to mask)
• No change to model weights, architecture, or API