Fix prior preservation weighting in advanced DreamBooth LoRA

[Subodh-17]

Description

Fixes the missing prior-preservation weighting split in the advanced DreamBooth LoRA training scripts.

When prior preservation is enabled, model_pred and target are split into instance and prior batches, but weighting was left unsplit. The prior loss therefore attempted to apply a weighting tensor covering both batches to the prior batch only, resulting in a shape mismatch.

This change splits weighting in the same way as model_pred and target and uses weighting_prior for the prior loss.

Changes

• Chunk weighting into instance and prior portions in the advanced Flux DreamBooth LoRA script.
• Chunk weighting into instance and prior portions in the advanced SDXL DreamBooth LoRA script.
• Use weighting_prior when calculating prior-preservation loss.

Scope

This PR is intentionally limited to the missing weighting chunking described in #11489.

Testing

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

The full prior-preservation training regression path could not be isolated into a dedicated test because the existing advanced Flux script has a separate pre-existing text_ids issue that occurs earlier in the training path. That unrelated issue is intentionally not modified by this PR.

Fixes #11489