Refactor EL models with standardized training loop

[ferzcam]

Move generic training logic to EmbeddingELModel base class, allowing subclasses to customize behavior through hooks instead of duplicating the entire training loop.

New customization hooks in EmbeddingELModel:

• get_negative_sampling_config(): Configure which GCIs need negatives
• generate_negatives(): Custom negative sampling strategy
• compute_loss(): Custom loss computation (e.g., MSE loss for ELBE)
• get_regularization_loss(): Add regularization from module
• get_optimizer(): Use different optimizer

API changes:

• epochs parameter moved from init() to train()
• Subclasses now only need to implement init_module() and override hooks as needed

Models simplified:

• ELBE: Only overrides compute_loss() for MSE loss
• ELEmbeddings: Uses base class entirely (has regularization)
• BoxSquaredEL: Uses base class entirely (has regularization)

Example models updated to use new API:

• ELBEPPI, ELBEGDA
• ELEmPPI, ELEmGDA
• BoxSquaredELPPI, BoxSquaredELGDA

Code reduction: ~640 lines removed, ~450 lines added (net -190 lines)