Qwen Image BS>1: remaining double-pass of encoder_hidden_states_mask after v4.1.2

## Context v4.1.2 fixed the padding corruption for Qwen Image BS>1 training — the core issue from #2647/#2648/#2655 is resolved. However there's one remaining edge case: `encoder_hidden_states_mask` gets passed twice to `self.attn()` in `DoubleStreamLayer`, causing a `TypeError` at step 0. ## Error ``` TypeError: Attention(...) got multiple values for keyword argument 'encoder_hidden_states_mask' ``` ## Cause In `helpers/models/qwen_image/transformer.py`: - Main `forward()` at line ~1204 stuffs the mask into `block_attention_kwargs` - `DoubleStreamLayer.forward()` also receives it as an explicit parameter - When `self.attn()` is called with both the explicit kwarg AND `**joint_attention_kwargs`, Python sees the kwarg twice ## Fix Remove line ~1204 (`block_attention_kwargs["encoder_hidden_states_mask"] = encoder_hidden_states_mask`). The mask already flows via the explicit parameter through the block call chain. ## Validated - ST 4.1.2, PRO 6000, Qwen Image v2.0, LoRA r32, BS=4 - With fix: clean validation samples at steps 25, 50, 75 - Happy to PR this — it's one line