// set of leading dimensions before doing the actual matmul.
  // That's what the code below does.
  // In particular, we populate out_lhs and out_rhs to have dimension structure:
  // - out_lhs: [broadcast(LHSBATCHDIMS, RHSBATCHDIMS)..., LHSROWS, LHSCOLS]
  // - out_rhs: [broadcast(LHSBATCHDIMS, RHSBATCHDIMS)..., RHSROWS, RHSCOLS]
  // To do this, we need to calculate those output shapes, which involves