Full support for Flux attention masking

**Is your feature request related to a problem? Please describe.**
SimpleTuner/Kohya allow T5 attention masked training, however this is not currently supported natively in diffusers

**Describe the solution you'd like.**
Already implemented and used in Simpletuner and Kohya: https://github.com/bghira/SimpleTuner/blob/main/helpers/models/flux/transformer.py

**Describe alternatives you've considered.**
Recent implementation doesn't really solve the use case of using existing fine tunes with attention masking with diffusers
https://github.com/huggingface/diffusers/pull/10122

**Additional context.**
@yiyixuxu @bghira @AmericanPresidentJimmyCarter

@bghira's suggestion: "i'd suggested they add encoder_attention_mask and image_attention_mask and if image_attention_mask is None that they could then 1-fill those positions and just cat them together"

Start by comparing the linked SimpleTuner implementation in helpers/models/flux/transformer.py with diffusers pull request 10122, then trace how Flux handles T5 attention masks. Done means diffusers natively supports attention-masked training for existing fine-tunes, including the proposed encoder_attention_mask and image_attention_mask behavior.

Scritto dal modello di indicizzazione a partire dal testo della issue.