Traversal logic in `evaluate` and other core routines should be shared

[eb8680]

There are currently a few different places in core library code that implement generic data structure traversals that really ought to be unified to ensure correctness and consistency of behavior:

• ops.semantics.evaluate
• internals.product_n.map_structure
• internals.unification.nested_type
• ops.syntax.Scoped.analyze
• ops.types.Term.__str__
• PyTree APIs in PyTorch and JAX via handlers.torch/handlers.jax

To start out, it should be relatively straightforward to refactor evaluate and map_structure to share their singledispatch-based traversal logic for non-Term data structures. Using the same traversal code in nested_type may be feasible as well, as should integrating with jax.tree.map for PyTree types.

[jfeser]

Some of these uses could become calls to evaluate if we implement #715 (e.g. Term.__str__, Scoped.analyze).