[mlir][tosa] Fix lowering of tosa.matmul with dynamic outputs (#72724)

The existing lowering of tosa.matmul will construct illegal tensor.empty
operations when the output type is more dynamic than the input types.

%0 = tosa.matmul %arg0, %arg1 : (tensor<1x1x8xf32>, tensor<1x8x1xf32>)
-> tensor<?x1x1xf32>

When constructing the tensor.empty operation, consult the output type
rather than the input types to decide whether a dimension is dynamic.