Apply a single, explicitly-selected LoRA adapter to ``module`` at scale 1.0. This is a fast, allocation-free replacement for the previous

Encodes the images into tokens and ids for FLUX pipeline.

AdaLayerNormZero.forward with LoRA applied to the inner ``linear``.

AdaLayerNormZeroSingle.forward with LoRA applied to the inner ``linear``.

FeedForward.forward routing each sub-module through ``lora_forward``. Only the output projection (``ff.net[2]``) is LoRA-wrapped in practice;