Do not coerce gradients to TensorVariable

[ricardoV94]

This could cause spurious disconnected errors, because the tensorified variable was not in the graph of the cost

---

📚 Documentation preview 📚: https://pytensor--1685.org.readthedocs.build/en/1685/

[ricardoV94]

How did this come up?

I was trying to use it in the loop autodiff example. Ended up not wanting it but fixed it anyway