[AutoDiff] Refine pullback generation logic for reference-based types

[asl]

Currently we are generating new adjoint buffers for many operations on class references that do look like projections. For example:

  /// Handle `upcast` instruction.
  ///   Original: y = upcast x
  ///    Adjoint: adj[x] += adj[y]
  ///             (assuming adj[x] and adj[y] have the same type)
  void visitUpcastInst(UpcastInst *ui) {
    auto *bb = ui->getParent();
    assert(ui->getOperand()->getType().isObject());
    assert(getRemappedTangentType(ui->getOperand()->getType()) ==
               getRemappedTangentType(ui->getType()) &&
           "Operand/result must have the same `TangentVector` type");
    switch (getTangentValueCategory(ui)) {
    case SILValueCategory::Object: {
      auto adj = getAdjointValue(bb, ui);
      addAdjointValue(bb, ui->getOperand(), adj, ui->getLoc());
      break;
    }
    case SILValueCategory::Address: {
      auto adjDest = getAdjointBuffer(bb, ui);
      addToAdjointBuffer(bb, ui->getOperand(), adjDest, ui->getLoc());
      builder.emitZeroIntoBuffer(ui->getLoc(), adjDest, IsNotInitialization);
      break;
    }
    }
  }

Note that here x and y represent essentially the same reference-counted objects, so they may share the adjoint buffer instead of allocating separate adjoint buffers for x and y respectfully. Accumulation and zeroing could be omitted as well.

[asl]

Tagging @rxwei @dan-zheng @BradLarson

[asl]

Same happens for unchecked_ref_cast Likely begin_borrow should be improved as well, as it is equivalent of strong_retain on reference-counted values.