Title

No LoRA fine-tuning example for RFD4-Proteina

Body

Is your feature request related to a problem? Please describe.

RFD4ProteinaPipeline ships with inference docs and a conversion script, but
examples/ has no training entry point for it. Fine-tuning the denoiser on a
set of structures is the main thing people want to do with the checkpoint after
generating from it, and today that means reimplementing the recipe from scratch:
the frozen autoencoder paired with the denoiser, the per-modality time
distribution, the product-space flow-matching loss over backbone coordinates and
local latents, and the noise draw that has to match what the sampler starts from.

Getting any of those wrong is quiet — training runs and the loss falls, but the
adapter does not move the model the way the released recipe does.

Describe the solution you'd like.

An examples/rfd4_proteina/ directory in the same shape as the other LoRA
examples: an 🤗 Accelerate training script, a README and a requirements.txt.
The script would follow the recipe the checkpoint was released with (rank 32 /
alpha 64 on the attention projections, the native per-modality time
distribution) and save adapters in the format
RFD4ProteinaPipeline.load_lora_weights already reads, so a fine-tune drops
straight back into the pipeline.

Describe alternatives you've considered.

Leaving fine-tuning to the model's own trainer outside Diffusers. That works,
but adapters produced there need converting before the pipeline can load them,
and it leaves the pipeline with no training path of its own.

Additional context.

Structures are read as annotated CIF files and featurized with
rfd4-proteina-prep, the
same preprocessing the pipeline's inference examples use.