Multi rotate cse

```
%32:3 = "enzymexla.multi_rotate"(%31) <{dimension = 2 : si32, left_amount = 2 : si32, right_amount = 0 : si32}> : (tensor<4x1520x1520xf64>) -> (tensor<4x1520x1520xf64>, tensor<4x1520x1520xf64>, tensor<4x1520x1520xf64>) loc(#loc705)
%33:4 = "enzymexla.multi_rotate"(%31) <{dimension = 2 : si32, left_amount = 3 : si32, right_amount = 0 : si32}> : (tensor<4x1520x1520xf64>) -> (tensor<4x1520x1520xf64>, tensor<4x1520x1520xf64>, tensor<4x1520x1520xf64>, tensor<4x1520x1520xf64>) loc(#loc707)
```

we shouldn't have both here @jumerckx ?

Start with the shown MLIR containing the two `enzymexla.multi_rotate` operations and trace the compiler's common-subexpression-elimination handling for this operation. Compare why the dimension and rotation amounts produce both results, and consider the issue done when redundant multi-rotate work is eliminated without changing the expected result arity.

Written by the indexing model from the issue text.