This is basically implemented, but I made it throw a NotImplementedError, since it looked like there were a few things missing. Notably, the typeFn variable is not used anywhere, even though other linalg operations use it for casting, and it currently uses tensor.empty which isn't really implemented. Maybe at first, instead of having tensor.empty, we can just make the user pass in the output, like what MLIR does and what we currently do for the other linalg operations. Also should write at least one testcase for batch_matmul.