Explicitly track all possible index access in stablehlo.while

https://github.com/EnzymeAD/Enzyme-JAX/pull/1668 generalizes indexing to `scale * i + offset`. However this wont work for IR like

```mlir
module {
func.func @main(%arg0: tensor<10xf64>) -> tensor<10xf64> {
%c = stablehlo.constant dense<1> : tensor
%c_0 = stablehlo.constant dense<0> : tensor
%c_1 = stablehlo.constant dense<10> : tensor
%c_2 = stablehlo.constant dense<3> : tensor
%cst = stablehlo.constant dense<0.000000e+00> : tensor<10xf64>
%c_3 = stablehlo.constant dense<2> : tensor
%0 = stablehlo.dynamic_slice %arg0, %c_3, sizes = [10] : (tensor<10xf64>, tensor) -> tensor<10xf64>
%1 = stablehlo.slice %0 [0:10:3] : (tensor<10xf64>) -> tensor<4xf64>
%2 = stablehlo.cosine %1 : tensor<4xf64>
%3 = stablehlo.dynamic_slice %arg0, %c_3, sizes = [10] : (tensor<10xf64>, tensor) -> tensor<10xf64>
%4 = stablehlo.slice %3 [0:10:3] : (tensor<10xf64>) -> tensor<4xf64>
%5 = stablehlo.sine %4 : tensor<4xf64>
%6:2 = stablehlo.while(%iterArg = %c_0, %iterArg_4 = %cst) : tensor, tensor<10xf64>
cond {
%7 = stablehlo.compare LT, %iterArg, %c_1 : (tensor, tensor) -> tensor
stablehlo.return %7 : tensor
} do {
%7 = stablehlo.add %c_2, %iterArg : tensor
%8 = stablehlo.convert %7 : (tensor) -> tensor
%9 = stablehlo.subtract %8, %c : tensor
%10 = stablehlo.divide %iterArg, %c_2 : tensor
%11 = stablehlo.dynamic_slice %5, %10, sizes = [1] : (tensor<4xf64>, tensor) -> tensor<1xf64>
%12 = stablehlo.dynamic_slice %2, %10, sizes = [1] : (tensor<4xf64>, tensor) -> tensor<1xf64>
%13 = stablehlo.subtract %12, %11 : tensor<1xf64>
%14 = stablehlo.dynamic_update_slice %iterArg_4, %13, %9 : (tensor<10xf64>, tensor<1xf64>, tensor) -> tensor<10xf64>
stablehlo.return %7, %14 : tensor, tensor<10xf64>
}
return %6#1 : tensor<10xf64>
}
}
}
```

If we instead track the full range of possible indices we can hoist even this loop