[AutoDiff] Add cloned curry thunks to generated function list.

[dan-zheng]

Add cloned curry thunks to generated function list so that they will be
deleted upon clean up if any diagnostics are emitted.

Resolves TF-918: crash due to invalid cloned curry thunk that was not deleted
during clean up.

---

Before:

tf-918.swift:1:44: error: function is not differentiable
 _ = gradient(at: Float(1), Float(2), in: (+) as @differentiable (Float, @nondiff Float) -> Float)
                                          ~^~
<unknown>:0: note: cannot convert a direct method reference to a '@differentiable' function; use an explicit closure instead
SIL verification failed: return value type does not match return type of function: functionResultType == instResultType
... # Crash after verification failure.

After:

tf-918.swift:1:44: error: function is not differentiable
 _ = gradient(at: Float(1), Float(2), in: (+) as @differentiable (Float, @nondiff Float) -> Float)
                                          ~^~
tf-918.swift:1:44: note: cannot convert a direct method reference to a '@differentiable' function; use an explicit closure instead
 _ = gradient(at: Float(1), Float(2), in: (+) as @differentiable (Float, @nondiff Float) -> Float)
                                           ^
# No crash.

[dan-zheng]

@swift-ci Please test tensorflow

[dan-zheng]

Note: function references to generated subset parameter thunks (for linear map and derivative functions) are not tracked. This is not a problem because all references to such thunks are in other generated functions, so cleanup already deletes all such references.

I decided not to track references to subset parameter thunks for now, since there's no problem.

[dan-zheng]

@swift-ci Please test tensorflow

[rxwei]

@swift-ci Please test tensorflow