Port of dequantize(): x = (indices + zero_point) * scale, output cast to bf16 (the cache's working dtype). indices must be UNPACKED (one element per array element) and scales/zeros are [..., dim/32].
indices
Port of dequantize(): x = (indices + zero_point) * scale, output cast to bf16 (the cache's working dtype).
indicesmust be UNPACKED (one element per array element) and scales/zeros are [..., dim/32].