first pretrained models support prototype

[maryapp]

Add Support for Pretrained DeepLabCut Models

issue: #141

Summary

This PR adds support for pretrained DeepLabCut models (e.g. SuperAnimal / Model Zoo) using the existing Model → PoseEstimationTask → PoseEstimation workflow.
No parallel schema is introduced; trained and pretrained models are handled as two modes of the same pipeline.

Key Changes

• PretrainedModel lookup

  • Registers pretrained models (pretrained_model_name, source, version, species, backbone_model_name, detector_name, default_params, weights_path, description).
  • Helpers: is_pretrained(...), add(...) for explicit registration.

• Model integration

  • Pretrained models detected via metadata (e.g. _pretrained_model_name in config).
  • Optional helper to insert pretrained models while validating against PretrainedModel.
  • No changes to existing Model schema.

• PoseEstimation inference

  • Extends make with two code paths:
    • pretrained → _do_pretrained_inference(...)
    • trained → existing logic via do_trained(...)
  • _do_pretrained_inference:
    • Uses PretrainedModel metadata.
    • Calls deeplabcut.video_inference_superanimal(...) when available.
    • Falls back to video_inference(...) where appropriate.
    • Filters kwargs via inspect.signature() to stay API-compatible.

Backward Compatibility

• Existing trained-model workflows are unchanged.
• No schema migration required.
• New behavior is enabled only when a model is explicitly configured as pretrained.

Testing - TODO:

• Pretrained workflow with a registered SuperAnimal model.
• Fallback to video_inference when video_inference_superanimal is unavailable.
• Parameter merging (defaults from PretrainedModel + task overrides).
• Validation for missing or unregistered pretrained models.