Implement motive decomposition network (MDN)

[Samrawitgebremaryam]

Summary

This PR implements the Motive Decomposition Network (MDN) for SubRep and adds unit tests for its forward-pass behavior.

The MDN maps context features to two context-conditioned outputs:

• strictly positive Dirichlet alpha parameters over objective weights
• bounded support geometry values for the admissible weight region

Changes

• add MotiveDecompositionNetwork in generator/mdn.py
• add a shared MLP trunk with separate distribution_head and support_head
• produce Dirichlet alpha parameters using Softplus + epsilon
• produce support geometry values using Sigmoid
• validate input shape with clear ValueErrors

Verification

• manual forward-pass checks completed for single, batched, and invalid inputs
• full test suite passes locally with:

python -m pytest -v