[autodiff] Should we add iterate-to-numeric-fixed-point implementation?

[sritchie]

I don't fully get this yet, but @axch defines iterate-to-numeric-fix for doing... well, "iterate to fixed point" style computations, like optimizers require. This is analogous to seq-limit, but works better with duals?

I don't get why we need this yet, given that we can definitely iterate to fixed point normally, like this: https://github.com/axch/dysvunctional-language/blob/b45150afa6f57b718521ac99aa3d3be1640993cb/vl/examples.scm#L116

The document @axch references explaining the situation is missing, but I'm sure it's just migrated to some other place:

  ;; For motivation (that is, why this function cares about bundles at all),
  ;; see ls/dvl/doc/num-fix.tex.
  ;;
  ;; This implementation follows Strategy 2 from that document, which
  ;; is to iterate the primal to convergence alone, and then use that
  ;; result to iterate the tangent to convergence.

SO! This might be a great idea.