Implement causal grouped-query attention without materializing repeated key or value heads. Queries have shape (batch, query heads, sequence, head width); keys and values use fewer or equal heads. Each contiguous group of G query heads shares one key/value head, where G is the query-head count divided by the key/value-head count.

Here h is a zero-based query-head index, and floor selects its shared key/value head. Q is one query head, K and V are its assigned key and value heads, D is head width, and the superscript T transposes K. The mask M is zero for keys at or before the query position and negative infinity for later keys. Softmax normalizes over key positions.

Return a PyTorch tensor with the same shape, dtype, and device as q.