[AutoDiff] Fix `partial_apply` substitution map logic. (#22240)
* [AutoDiff] Fix `partial_apply` substitution map logic.
Propagate derivative's substitution map in `reapplyFunctionConversion` for
`partial_apply` instructions. When reapplying a `partial_apply` instruction,
the associated derivative function's substitution map should be used, not
the original function's. This is because the derivative function may have
more generic requirements than the original function.
Revamp `test/TensorFlowRuntime/model_autodiff_runtime.swift` to use imported
TensorFlow high-level APIs (layers and optimizers).
Resolves SR-9799.
* Clean up substitution map logic.
- Create `assocFn` without generic environment.
- Clean up substitution map logic.
- Compute `partial_apply` substitution map using derivative's generic signature.
Address comments from @slavapestov.
