[AutoDiff] Fix control flow AD when OME is skipped.

[dan-zheng]

-enable-strip-ownership-after-serialization is now true by default,
skipping ownership model eliminator after differentiation.

This is problematic for control flow differentiation. VJPEmitter creates
trampoline bbs, which create predecessor enum values and immediately branch to
a destination bb. Trampoline bbs share the argument of their destination bb,
which may have @guaranteed ownership kind. @guaranteed values need a single
lifetime-ending use, e.g. an end_borrow.

Previously, VJPEmitter::trampolinedGuaranteedPhiArguments tracked all
@guaranteed trampoline/destination bb argument pairs and emitted an
end_borrow for the trampoline bb argument after the cloned end_borrow for
the destination bb argument. However, this does not work when ownership model
eliminator is skipped: mandatory inlining invokes mergeBasicBlocks on VJPs,
which causes the end_borrow instructions to be redundant.

// Generated by differentiation:
bb1(%6 : @guaranteed $Tracked<Float>):            // Preds: bb0
  %7 = enum $_AD__$s1e15enum_notactive2y23DifferentiationUnittest7TrackedVySfGAA4EnumO_AFtF_bb1__Pred__src_0_wrt_1, #_AD
__$s1e15enum_notactive2y23DifferentiationUnittest7TrackedVySfGAA4EnumO_AFtF_bb1__Pred__src_0_wrt_1.bb0!enumelt.1, %4 : $
_AD__$s1e15enum_notactive2y23DifferentiationUnittest7TrackedVySfGAA4EnumO_AFtF_bb0__PB__src_0_wrt_1 // user: %8
  br bb2(%6 : $Tracked<Float>, %7 : $_AD__$s1e15enum_notactive2y23DifferentiationUnittest7TrackedVySfGAA4EnumO_AFtF_bb1_
_Pred__src_0_wrt_1) // id: %8
// %9                                             // user: %11
// %10                                            // user: %13
bb2(%9 : @guaranteed $Tracked<Float>, %10 : $_AD__$s1e15enum_notactive2y23DifferentiationUnittest7TrackedVySfGAA4EnumO_A
FtF_bb1__Pred__src_0_wrt_1): // Preds: bb1
  end_borrow %9 : $Tracked<Float>                 // id: %11
  end_borrow %6 : $Tracked<Float>                 // id: %12
  ...

// After mandatory inlining, which calls `mergeBasicBlocks`:
bb1(%6 : @guaranteed $Tracked<Float>):            // Preds: bb0
  %7 = enum $_AD__$s1e15enum_notactive2y23DifferentiationUnittest7TrackedVySfGAA4EnumO_AFtF_bb1__Pred__src_0_wrt_1, #_AD__$s1e15enum_notactive2y23DifferentiationUnittest7TrackedVySfGAA4EnumO_AFtF_bb1__Pred__src_0_wrt_1.bb0!enumelt.1, %4 : $_AD__$s1e15enum_notactive2y23DifferentiationUnittest7TrackedVySfGAA4EnumO_AFtF_bb0__PB__src_0_wrt_1 // user: %10
  end_borrow %6 : $Tracked<Float>                 // id: %8
  end_borrow %6 : $Tracked<Float>                 // id: %9
  ...

Function: 'AD__$s1e15enum_notactive2y23DifferentiationUnittest7TrackedVySfGAA4EnumO_AFtF__vjp_src_0_wrt_1'
Found over consume?!
Value: %13 = argument of bb2 : $(Tracked<Float>, Tracked<Float>) // users: %15, %17, %18
User:   end_borrow %13 : $(Tracked<Float>, Tracked<Float>) // id: %17
Block: bb2

Now, VJPEmitter preemptively calls mergeBasicBlocks, which merges trampoline
bbs with their destinations. @guaranteed bb arguments are merged and no longer
duplicated, so there is no need to specially emit end_borrow for them.

Consequences:

• VJP functions now have simplified CFGs.
• All VJP trampoline bbs are merged with their destinations. Consider rewriting
  the transform to generate the merged destination bbs in the first place.

Resolves TF-990.

[dan-zheng]

@swift-ci Please clean test tensorflow

[dan-zheng]

I created a PR to document this VJP generation change.
Happy to discuss in detail later, merging now to unblock progress.