fix: remove incorrect text_ids concatenation in Flux DreamBooth LoRA training

[Subodh-17]

Description

Fixes #10722.

The Flux DreamBooth LoRA training scripts still concatenate text_ids with class_text_ids when prior preservation is enabled.

text_ids is now a 2D tensor containing sequence positional IDs. FluxTransformer2DModel broadcasts this tensor internally across the batch, so concatenating instance and class text_ids along dimension 0 incorrectly doubles the sequence dimension.

This causes a sequence-length mismatch later when the transformer constructs the rotary embeddings.

Changes

Remove the incorrect text_ids concatenation from the static prior-preservation path in:

• examples/dreambooth/train_dreambooth_lora_flux.py
• examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py

The resulting text_ids shape matches the format expected by FluxTransformer2DModel and the existing dynamic prompt path.

Background

PR #9653 changed Flux text_ids from a batched tensor to a 2D tensor because the transformer handles batch broadcasting internally. The prior-preservation static caching path retained the older concatenation logic.

Testing

• python -m py_compile examples/dreambooth/train_dreambooth_lora_flux.py examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py
• ruff check examples/dreambooth/train_dreambooth_lora_flux.py examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py
• git diff --check

The repository currently does not provide an isolated prior-preservation regression test for these training scripts, so no new end-to-end test is included in this minimal fix.

Fixes #10722