Add ring-shared weights and same depth padding

[YugnatD]

Summary

• Add optional ring-shared weights via share_neighbors for Conv2d and Conv3d.

• Add Conv3d depth_padding="same" to preserve the depth dimension with centered temporal kernels.
• Materialize dense kernels during forward so the existing convolution implementation remains unchanged.

Tests

• PYTHONPATH=src python3 -m pytest (134 passed)