Implement ModularDNN model

[stefan-apollo]

Description

This PR implements the ModularDNN model used for our toy experiments. It is based off MLP but not loaded from a trained model but instead initialised randomly according to some conditions.

The experiments folder does contain a "notebook" which reproduces Jake's main experiments.

This branch also implements the "neuron" basis.

Replaces #227 (Jake's branch).

Based off feature/support_different_bases_and_edges but the diff looks fine?

Tested

Testing

• Test ModularDNN build graph
• Test invariance to rotate_final_layer (though currently broken, see issue)
• Test a linear model (Identity activation) gives a trivial graph
• build graph for neuron basis (but don't check Lambdas because they don't exist)

Breaking changes

No, I don't think so.

[stefan-apollo]

Jake suggests renaming to CustomMLP because he wants to also try non-modular examples

[danbraunai-apollo]

The new code that runs the block diagonal experiments within mlp_rib_build is able to replicate the code from before my commits 19206e2 by running a build on block_diagonal.yaml and changing the truncation_threshold to 1e-6.