circuit-tracer integration for CrossLayerTranscoders

[jiito]

Changes

1. Added ModelConverter and concrete CircuitTracerConverter that is used to turn a CLTModule into circuit tracer compatible format.
2. Added callbacks for hugging face upload and model conversion.
3. Added integration test to verify model conversion + loading in circuit-tracer.

Questions

1. When is the best time to save the circuit-tracer compatible model? Which callback to attach it to in the checkpoint? I have made this configurable with config arguments
2. Do we have to adapt how we save the non-linearity threshold? The circuit tracer expects a [1, d_features] tensor for each layer's non-linearity. Right now, I am using model.nonlinearity.theta[:, layer_idx, :] but this doesn't account for the bandwidth...
3. Any better suggestions for the ModelConverter interface? This is maybe better framed as a ModuleSaver? I am planning for using this same interface for delphi and other integrations.

Verification

• integration test with an untrained model
• tested the ModelConversionCallback manually
• Functional E2E test with training (circuit-tracer.yaml) and loading the saved model.
• Double check non-lin settings

Resolves #4

Future Work

• Because circuit-tracer only supports relu and jump relu activation functions, additional work needs to be done to address compatibility with topK and other types of CLTs.
• Take a look at composability of yaml config files to greatly simplify / modularize aspects of the training process.