Embedding scaling for TrainableTokensModel in XLoraEmbeddingLayer

[sambhavnoobcoder]

System Info

peft: 0.17.1
transformers: 4.56.2

Who can help?

@benjaminbossian

This issue tracks the fix needed for XLoraEmbeddingLayer, corresponding to the changes in #2809 for LORA.
The parent issue addresses LORA updates, but a similar adjustment is required for XLoRA to ensure consistent behavior.

Creating this issue for tracking purposes — will push a PR for the same shortly.

Reproduction

#2809

Expected behavior

Adapter should distinguish between vanilla torch.nn.Embedding layer and custom variants.