Fix double scaling when exporting Gemma embeddings

[ybochkov]

PartlyFrozenEmbeddings.to_embeddings() previously rebuilt the embedding table by calling self.forward() for every token. Gemma's specialized wrapper applies embed_scale in forward(), so the exported checkpoint contained already-scaled vectors. Gemma then applied the same scale again at inference time after loading the checkpoint.

Reconstruct the standard embedding table directly from the frozen and trainable parameter partitions instead. Preserve the source padding index and dtype while rebuilding the layer.

Add regression tests that distinguish runtime Gemma scaling from persisted raw weights and verify that ordinary embedding reconstruction retains weights and metadata.