Enable taking derivatives with respect to multiple variables simultaneously

[thomasahle]

Background
Our grad mechanism currently only handles differentiation w.r.t. a single Variable. For certain use cases (like computing Jacobians or Hessians, or typical backprop on a large network), we might want partial derivatives with respect to all parameters in one pass.

Potential Approaches

• Extend the Derivative class to store a tuple/list of variables.
• Reuse single-variable derivatives but memoize or reuse computations so we don’t reevaluate sub-expressions multiple times.
• Exploit the isomorphic hashing approach to avoid redundant computations (caching repeated sub-graphs).

Additional Context

• This is important for scenarios like second-order optimization (Hessian-based methods) or certain advanced autodiff use cases.