feat(config): opt in to typed nested dataclass construction

[hasuoshenyun]

What does this PR do ?

Allow dataclass fields to explicitly opt into typed nested YAML construction at the existing ConfigNode boundary.

Changelog

• Support field(metadata={"instantiate": True}) for unambiguous typed nested dataclasses and homogeneous sequences.
• Opt in an entire supported typed subtree: plain nested dataclass fields do not need additional metadata, and homogeneous list/tuple fields follow their annotations.
• Preserve existing list/dictionary behavior for unmarked fields on the explicitly instantiated target, outside opted-in subtrees.
• Do not coerce runtime overrides or resolve their contents as environment variables.
• Preserve construction behavior when runtime type hints cannot resolve TYPE_CHECKING-only references. Resolution is per class, not per field: a NameError or TypeError disables annotation-based field conversion for that class, so even marked fields can retain plain dictionaries. Existing explicit nested _target_ behavior is unchanged.
• Add regression tests independent of QAT imports, including three-level construction, unmarked sibling preservation, repeated instantiation, runtime override identity, and unchanged source configuration.

This is a configuration-construction convenience, not a training or quantization algorithm change. Explicit nested _target_ configurations remain supported; this opt-in lets annotations supply the nested dataclass types instead of repeating class paths in YAML.

Before your PR is "Ready for review"

• Read the contributor guidelines and signed off commits.
• Added focused compatibility tests.
• Documented the explicit opt-in contract.

Additional Information

This PR is now independent of #3902 and rebased onto upstream main at 44cf34834. Its diff contains only nemo_automodel/components/config/loader.py and tests/unit_tests/config/test_loader.py; no expert LoRA or QAT implementation changes are included. The two config commits are 718ab3a00 and 10f9d6c6d. Rebase range-diff confirms both patches are unchanged from their previously published versions.

The implementation deliberately does not enable type coercion globally: differential checks against the original loader confirm unchanged list/dict construction for ordinary dataclasses.

Validation rerun on the independent head 10f9d6c6de23444bd08f70c48e30e3dc0ccde5b2: 97 passed across tests/unit_tests/config/, including the three-level subtree regression and existing unresolved-hint fallback tests. Loader Ruff lint/format and git diff --check pass. Upstream CI for this SHA still requires maintainer authorization; local results are not a substitute for upstream CI.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[hasuoshenyun]

Hi @akoumpa, this PR is now ready for review and independent of #3902, rebased onto main. The diff contains only the config loader and its tests. The recursive opt-in/fallback contract is documented, and all 97 config tests pass locally on head 10f9d6c. Could you please review and authorize CI for this SHA when convenient? Thank you!