feat(image): add MetaFormer encoder integration

[khaivandangusf2210]

feat(image): add MetaFormer encoder integration

Summary
Adds a new image encoder based on the MetaFormer family of architectures to Ludwig, enabling users to leverage a flexible token-mixing backbone for image feature extraction. Provides:

• Encoder implementation: ludwig/encoders/image/metaformer.py
• Schema / config validation: ludwig/schema/encoders/image/metaformer.py
• Model variants + stacking helpers (metaformer_integration/)
• Tests: tests/ludwig/encoders/test_metaformer_encoder.py
• Minimal reproducible image dataset (test_data/) for fast CI validation

Motivation
MetaFormer-style backbones (e.g. ConvNeXt-like and token-mixing generalizations) offer competitive performance and architectural simplicity. This integration broadens Ludwig’s catalog of vision encoders and facilitates experimentation and benchmarking against existing CNN and ViT-style encoders.

Implementation Details

• Introduces MetaFormerEncoder class implementing the Ludwig image encoder interface.
• Supports selectable variants (defined in metaformer_integration/metaformer_models.py) and a stacked CNN + MetaFormer hybrid (metaformer_stacked_cnn.py).
• Registered in ludwig/encoders/image/init.py and corresponding schema init to enable config-based discovery.
• Configuration validation via dedicated schema class including fields: variant, pretrained (if applicable later), dropout, embedding size, and internal layer settings.
• Keeps external dependencies minimal; no new pip requirements added (relies on existing torch stack).

New Files

• ludwig/encoders/image/metaformer.py
• ludwig/schema/encoders/image/metaformer.py
• metaformer_integration/init.py
• metaformer_integration/metaformer_models.py
• metaformer_integration/metaformer_stacked_cnn.py
• tests/ludwig/encoders/test_metaformer_encoder.py
• test_data/ (tiny MNIST subset + few class images) for deterministic encoder test

Modified Files

• ludwig/encoders/image/init.py (registration)
• ludwig/schema/encoders/image/init.py (schema registration)
• ludwig/encoders/image/base.py + ludwig/schema/encoders/image/base.py (minor integration hooks if needed)
• ludwig/utils/tokenizers.py (unrelated minor touch if formatting import ordering — auto-format)

Testing
Test file tests/ludwig/encoders/test_metaformer_encoder.py covers:

• Construction from declarative config
• Forward pass shape consistency
• Basic training step (gradient flows)
• Variant parameterization logic (at least one additional variant)
  The included miniature dataset ensures test reproducibility without external downloads.

Reproducibility
A small curated sample (test_data/) enables running:
pytest tests/ludwig/encoders/test_metaformer_encoder.py -k metaformer
(Works offline and fast.)

Usage Example
YAML snippet:
input_features:

• name: image_path
  type: image
  encoder:
  type: metaformer
  variant: base # see metaformer_models.py for available variants
  dropout: 0.1
  output_features:
• name: label
  type: category

trainer:
epochs: 2

Backward Compatibility
No breaking changes. New encoder is opt-in. Existing image encoders unaffected.

Performance / Future Extensions

• Potential future support: loading external pretrained weights, mixed-precision tuning, more variants.
• If repository size concerns arise, the embedded test_data/ could be replaced with synthetic generation or hosted externally.

Issue Reference
If an issue exists, add a line such as:
Closes: #ISSUE_NUMBER
(Currently no explicit issue number provided.)

Checklist

• Encoder implementation
• Schema + validation
• Unit / integration style test
• Reproducible minimal data
• Documentation page update (can be a follow-up PR)
• Optional pretrained weight integration (future)

Let me know if you would like the docs updated in this PR; otherwise this focuses on the core encoder integration and tests.