Insert gradient reduction during pass expansion when multiple consumers use the same input

[elliottslaughter]

In the backward pass, a value used multiple times by different nodes results in gradients that must be summed together to produce the final result. This PR adds a new GradientReduction node type to implement this, along with a subgradient_id_t on DynamicValueAttrs to track the distinct sub-gradients that contribute to the final result, and then updates pass expansion to detect the duplicate values in the backward pass and insert GradientReductions for them.

---

This change is