Only apply SIMD vectorization to tensors that are secret

Now that we have `SecretnessAnalysis`, we could update the `--heir-simd-vectorizer` to only apply its transformations to tensors that are indeed `secret`.

It might be a good idea to make the implementation a bit more generic, so that one can specify an analysis and only have it apply to values where the analysis is true (e.g., for advanced scenarios where one is mixing SIMD arithmetic FHE and non-SIMD boolean FHE).