Fix fallback for backbone-only EAT checkpoints

[nkundiushuti]

When a checkpoint has no classifier and no label mapping, automatically switch to embedding mode for models that support return_features_only.
More details:

• list_model_layers() loads esp_aves2_eat_* with return_features_only=False, but those checkpoints are backbone-only (no classifier / label map), triggering num_classes must be > 0.
• load_model() now detects “checkpoint present but no classifier found” and automatically switches to return_features_only=True when supported.
• Added a unit test covering backbone-only checkpoint fallback.

[Copilot]

Pull request overview

Adjusts model loading to gracefully handle backbone-only EAT checkpoints (no classifier head and no label mapping) by automatically falling back to embedding extraction mode when the model supports return_features_only.

Changes:

• Update _load_from_modelspec() to enable return_features_only=True when a checkpoint exists but num_classes can’t be inferred and the model supports feature-only mode.
• Add a unit test covering the “checkpoint present but no classifier detected” fallback behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
avex/models/utils/load.py	Broadens the fallback logic so backbone-only checkpoints trigger embedding-mode loading when supported.
tests/unittests/test_api_load.py	Adds a unit test asserting that backbone-only checkpoints cause automatic fallback to return_features_only.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.