feat(backend): add support for xlabs Flux LoRA format

[Pfannkuchensack]

Summary

Add support for loading Flux LoRA models in the xlabs format.

The xlabs format uses a different key structure than other Flux LoRA formats:

• double_blocks.X.processor.{qkv|proj}_lora{1|2}.{down|up}.weight

Where:

• lora1 → image attention stream (img_attn)
• lora2 → text attention stream (txt_attn)
• qkv → query/key/value projection
• proj → output projection

Changes

• Add FluxLoRAFormat.XLabs enum value to taxonomy
• Add flux_xlabs_lora_conversion_utils.py with format detection and conversion
• Update formats.py to include xlabs in the detection cascade
• Update lora.py loader to handle xlabs format
• Update model probe in configs/lora.py to accept recognized Flux LoRA formats (fixes installation of xlabs LoRAs)
• Add unit tests for xlabs format detection and conversion

Related Issues / Discussions

Adds support for xlabs-format Flux LoRAs which were previously rejected with "model does not match LyCORIS LoRA heuristics".

Example LoRA using this format: Flux Realism LoRA

QA Instructions

1. Download an xlabs-format Flux LoRA (e.g., flux-RealismLora.safetensors from XLabs-AI)
2. Install the LoRA via Model Manager → Import Models
3. Verify it's recognized as a FLUX LoRA
4. Use the LoRA in a Flux generation
5. Verify the LoRA effect is applied to the output

Merge Plan

Standard merge, no special considerations.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[lstein]

I tested four different XLabs-format Flux LoRAs downloaded from the HuggingFace XLabs AI site. All were correctly recognized as Flux LoRAs and had the expected influence on generated images. In addition, I compared an XLabs format LoRA (disney_lora) to one that XLabs preconverted to standard format (disney_lora_comfy_converted), and both generated exactly the same image, confirming that the PR's conversion code is working properly.