Add multimodal model training support in WeightsLab

[guillaume-byte]

Summary

Enable multimodal model training workflows in WeightsLab (e.g., mixed modality inputs such as image+text, or other multimodal combinations).

Scope

• Add/extend data interfaces for multimodal batches.
• Ensure trainer lifecycle supports modality-specific preprocessing/forward/metrics.
• Support evaluation for multimodal models.
• Provide at least one baseline multimodal config/example.

Acceptance Criteria

• At least one multimodal training scenario runs successfully end-to-end.
• Data pipeline and trainer correctly process mixed modality inputs.
• Evaluation path works for multimodal models.
• Minimal documentation/example is added.

Dependency

This issue is blocked by #198.