When broadcasting, each 2D plane of src0 is matched with multiple 2D planes of src1. Planes of src0 need to be copied and/or de-quantized only once per multiple GEMM operations. A more natural way to handle this is to create outer loop over src0 and do copying and de-quantisation there.

Previously, de-quantization was performed before each GEMM. Now it is moved to an outer loop.

There is still duplication in case of broadcasting over dimension 3. Handling that properly would require a more substantial change, i.e. storing 3D instead of 2D slices of src0 in VRAM. I don't know of a case where broadcasting over dimension 3 is currently used, so I leave it for the future. Nevertheless, it produces correct results even in this case.

In case of matrix-vector multiplication, de-quantization is done repeatedly because de-quantized data is not stored in RAM.

Tested in isolation and with models that use GQA.