feat(differentiation): GPU default ad_backend AutoZygote -> AutoMooncake

[ocots]

Swaps the GPU default :ad_backend from AutoZygote() to AutoMooncake(), based on an H200 measurement of every reverse-mode DI candidate end to end: CTFlows PR #353.

Why

AutoMooncake cleared the entire matrix, including the real Flow(ocp) Hamiltonian call on a CuArray with a numerically correct result. AutoZygote turned out to fail unexpectedly on a non-mutating hamiltonian_gradient device call (KernelException) — using the exact same Hamiltonian formula that succeeds through a different call shape elsewhere. Its device behaviour is call-context dependent, not a stable pass/fail — a materially weaker guarantee than what a shipped default should offer. AutoZygote remains selectable explicitly; it's simply no longer the default.

What changed

• src/Differentiation/default.jl: __ad_backend(::Type{GPU}) = AutoMooncake().
• src/Differentiation/differentiation_interface.jl: the :ad_backend option's description is now P-dependent instead of one static string shared by CPU/GPU. describe(:di, registry) now shows a GPU-specific message — verified manually:

  ad_backend (backend, ad)::ADTypes.AbstractADType (default: ADTypes.AutoMooncake() [computed])
     description: DifferentiationInterface.jl backend for GPU execution. Default: AutoMooncake(),
     validated end-to-end on CuArray including through a mutating in-place RHS. AutoForwardDiff()
     does not work on GPU (scalar-indexes a CuArray). AutoZygote() can be selected explicitly; it
     was found unreliable in some device call contexts (an unexpected KernelException was
     observed on a non-mutating call).
  

• No new dependency — ADTypes.AutoMooncake() is a marker type from ADTypes (already a hard dep), same pattern AutoZygote() used before it.
• Tests: GPU default assertions updated; new assertions pin the GPU-specific description text and confirm CPU's is unaffected.
• Version 0.28.1-beta → 0.28.2-beta; CHANGELOG.md/BREAKING.md entries (non-breaking: same option name, override mechanism, computed=true flag — only the computed default value changes, on the opt-in GPU path).

Verification

Targeted suite/differentiation: 110/110. Full suite: 4758/4758.

🤖 Generated with Claude Code