feat(registry): add public architecture registration and entry-point discovery

[pstjohn]

Adds public out-of-tree architecture registration.

Custom model architectures (e.g. multimodal VLMs built by decorating a Hub-resolved text-only model) must register with AutoModel's model registry before construction. Today that requires importing the private nemo_automodel._transformers.registry.ModelRegistry and calling register() in every process that resolves architectures — including isolated Ray workers and CLI subprocesses where import-order dependencies are fragile.

_ModelRegistry now scans nemo_automodel.architectures entry points on first access, same shape as vLLM's vllm.general_plugins. Entries are arch_name = module.path:ClassName and are added to the lazy mapping without importing the target module; import happens on first architecture resolution, exactly like the built-in MODEL_ARCH_MAPPING entries.

We also export a public register_architecture():

from nemo_automodel import register_architecture
register_architecture("MyArch", MyModelClass)

This just delegates to the same _ModelRegistry.register, so explicit and entry-point-discovered registrations share one code path.

With entry-point discovery, you can now declare their architecture once in your pyproject.toml and every process discovers it automatically:

[project.entry-points."nemo_automodel.architectures"]
MyArch = "my_package.model:MyModelClass"

This uses Python's standard importlib.metadata.entry_points() (PEP 621).

[github-actions]

🌿 Preview your docs: https://nvidia-preview-preview-ae0013712db0.docs.buildwithfern.com/nemo/automodel

[HuiyingLi]

Wow thank you @pstjohn for adding it!

[jgerh]

Completed tech pubs review of docs/model-coverage/overview.mdx and provided some copyedits and suggested text revisions.

[jgerh]

Completed tech pubs review of docs/model-coverage/overview.mdx and provided copyedits and duplicate text deletions.

[pstjohn]

/ok to test ae00137