[bghira]

Accelerate's generic regional compiler also wraps MOSS auxiliary modules with different fixed parameter shapes, including the shared block and output-head modulation layers. These specializations can exhaust Dynamo's shared wrapper cache before image-size variation is considered.

Add an opt-in model-foundation context that uses Diffusers' existing compile_repeated_blocks() and _repeated_blocks declaration. MOSS enables this policy, compiling transformer blocks in place while suppressing Accelerate's automatic auxiliary-region compilation during preparation. Compiler settings are restored even if preparation fails. Training and inference preparation use the same hook; other model families retain their current policy.

Validation:

• .venv/bin/python -m unittest -v tests.test_moss_image tests.test_trainer: 153 tests pass, including output/gradient parity, parameter identity, compilation scope, plugin restoration, and training/inference wiring.
• Two H100s, tiny MOSS fixture (width 32, ten blocks): actual trainer preparation and three BF16 LoRA optimizer steps pass with both DDP and FSDP2, using ten transformer blocks, gradient checkpointing, dynamic image/caption sizes, and Inductor with fullgraph enabled. Auxiliary modulation layers remain eager.
• The two-H100 DDP check also passes with recompile_limit=8 and fail_on_recompile_limit_hit=True, so hitting the limit cannot silently fall back to eager execution.
• Full suite: .venv/bin/python -m unittest -v -f completed 7,052 tests successfully (292 skipped).
• All pre-commit hooks and git diff --check pass.

Closes #3286.