Write a linear approximations script

[DavidUdell]

Remaining todos here are:

1. implement credit assignment to error tensors,
2. clean up old gradient detaching code, which isn't used in this implementation, and
3. write code taking in sublayer sparse autoencoders and outputting intra-block causal graphs.

[DavidUdell]

Node level stuff is done. All that remains for the gradient-based implementation is the edge-level stuff.

[DavidUdell]

Edge level stuff done, excluding MLP-out and attn-out.