[lora] fix non-diffusers lora key handling for flux2

[sayakpaul]

What does this PR do?

Fix #13107

Cc: @mshehzadkhan. Could you give this a try please?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[mshehzadkhan]

I’ve tested a bunch of Klein LoRAs for each model from CivitAI, and they all seem to be loading correctly now.

Thanks a lot @sayakpaul for the quick fix 🙌

[dg845]

Thanks! If I understand correctly, the changes can be summarized as follows:

1. Support LoRAs which name the A and B matrices lora_down/lora_up instead of lora_A/lora_B respectively.
2. Support LoRAs which only target a subset of the transformer block weights (e.g. a subset of the attn QKV projections, or attn out projections, etc.)
3. Support LoRAs which target weights outside of the transformer blocks, as defined by extra_mappings.