try different model architectures

[falcondai]

RGB and D due to their acquisition are not naturally aligned

TODO

• [x] concat RGB+D
• [x] process RGB and D independently and merge near the last layer
• [x] attach a depth prediction head at an intermediate layer: RGB -> h [-> D; -> grasp valuation]

[falcondai]

having depth regression as an extra supervision during training improves the accuracy

concat RGB+D -> 92% acc late concat -> 89% RGB only -> 90% depth only -> 73% RGB only with extra depth supervision -> 91%