[AutoDiff] Start supporting loadable types with address-only tangents.

[dan-zheng]

Previously, PullbackEmitter assumed that original values' value category
matches their TangentVector types' value category. This was problematic
for loadable types with address-only TangentVector types.

Now, PullbackEmitter starts to support differentiation of loadable types with
address-only TangentVector types. This patch focuses on supporting and testing
class types, more support can be added incrementally.

Resolves TF-1149.
Unblocks future _Differentiation module changes like TF-1267.

---

Example:

import _Differentiation

// Class<T>: loadable type
// Class<T>.TangentVector: address-only struct type

class Class<T: Differentiable>: Differentiable {
  @differentiable
  var stored: T

  init(_ stored: T) {
    self.stored = stored
  }

  // struct TangentVector: Differentiable, AdditiveArithmetic {
  //  var stored: T.TangentVector
  // }
}

Before: cryptic error due to artificial limitations (lack of compiler support).

$ swift tf-1149.swift
tf-1149.swift:7:4: error: function is not differentiable
  @differentiable
  ~^~~~~~~~~~~~~~
tf-1149.swift:8:7: note: when differentiating this function definition
  var stored: T
      ^
tf-1149.swift:8:7: note: cannot yet differentiate value whose type 'Class<T>' has a compile-time known size, but whose 'TangentVector' contains stored properties of unknown size; consider modifying 'Class<τ_0_0>.TangentVector' to use fewer generic parameters in stored properties
  var stored: T
      ^

After: no error.

[dan-zheng]

@swift-ci Please smoke test

[rxwei]

This is not always true. Address only original types can totally have an non-address-only tangent, simplest example:

struct Foo<T>: Differentiable {
    var a: Float
    @noDerivative var b: T
}

Could you add a FIXME here with an appropriate bug number?

[dan-zheng]

Address only original types can totally have an non-address-only tangent

This case is actually currently supported today. The differentiation transform always creates adjoint buffers for address-typed original values, regardless of whether the tangent type is loadable or address-only.

Offline, you pointed out that we should make "tangent value category" depend solely on whether the tangent type is loadable or address-only. For loadable tangent types, using symbolic adjoint values instead of concrete adjoint buffers would be more efficient.

I filed SR-13077 to track this improvement and left a comment in the code.

[dan-zheng]

@swift-ci Please smoke test