Add Equivariant LoRA support to PoseGNN model

[luispintoc]

Summary

I’d like to extend PoseGNN with Equivariant LoRA (ELoRA) support to improve parameter efficiency while preserving geometric equivariance.

Motivation

• LoRA allows efficient fine-tuning by decomposing weight updates into low-rank matrices, specially for small datasets.
• This could make fine-tuning faster and less memory-intensive without breaking pose equivariance.

References

• LoRA: Low-Rank Adaptation of Large Language Models (Hu et al., 2021)
• ELoRA: Low-Rank Adaptation for Equivariant GNNs (Wang et al., 2025)

[neumannrf]

@davebraines I tried to tag Fabian for his awareness, but his GitHub tag doesn't show up here. 🤷‍♂️