[QUESTION] Why is Grouped-Query Attention (GQA) not supported in CrossAttention?

## Question
When trying to enable Grouped-Query Attention (GQA) in the cross-attention module, the following check in the current `CrossAttention` class prevents it:

```python
if self.config.num_query_groups != self.config.num_attention_heads:
raise ValueError("Group query attention is not currently supported in cross attention.")
assert self.query_projection_size == self.kv_projection_size
```

This effectively disables GQA whenever num_query_groups < num_attention_heads, even though GQA is supported in self-attention.

I would like to understand the reasoning behind this limitation:

- Is there a technical or architectural constraint that makes GQA incompatible with cross-attention?

- Are there known issues with masking, KV-cache handling, or parallelism that block support?

From a user perspective, GQA could significantly reduce memory usage and KV-cache size for encoder–decoder or retrieval-augmented models where cross-attention dominates cost.

It would be wonderful if future versions could support this feature.