[AutoDiff] Fix autodiff_function instruction worklist.

[dan-zheng]

Fix autodiff_function instruction worklist registration/invalidation.

• In VJPEmitter, add cloned autodiff_function instructions in VJP
  to worklist.
• If autodiff_function instructions are deleted in
  foldAutoDiffFunctionExtraction, also remove their worklist occurrences.

Resolves TF-515.

---

Writing a reproducer test is difficult because the nature of the bug is ad-hoc.
Verified that swift test works for tensorflow/swift-models@f9687bf.

[rxwei]

In Differentiation::run(), all autodiff_function instructions in a module are pushed to the work list. Could you please explain why it is necessary to push more in VJPEmitter::visitAutoDiffFunctionInst?

[dan-zheng]

In Differentiation::run(), all autodiff_function instructions in a module are pushed to the work list. Could you please explain why it is necessary to push more in VJPEmitter::visitAutoDiffFunctionInst?

If an original function has an autodiff_function instruction, it will be found and pushed to the worklist.
However, when VJPEmitter clones the original function into a VJP function, the newly cloned autodiff_function instruction needs to be manually added to the worklist.

Otherwise, the cloned autodiff_function instruction in the VJP won't be processed, leading to verification errors:

Assertion `!getAssociatedFunctions().empty() && "No associated functions. Maybe " "the differentiation pass has not run?"'

[dan-zheng]

@swift-ci Please test tensorflow

[dan-zheng]

@swift-ci Please clean test tensorflow

[rxwei]

It's a little unfortunate that while the cloned autodiff_function is an exact copy of the original, it has to go through emitAssociatiedFunctionReference twice.

[dan-zheng]

It's a little unfortunate that while the cloned autodiff_function is an exact copy of the original, it has to go through emitAssociatiedFunctionReference twice.

Yeah, that is true.