Add CIFAR-10 Classification Results

[siddk]

Add pipeline for CIFAR-10 Classification, as well as smoothing out training loop kinks. Additionally updates some of the exposition in the literate s4.py implementation.

Making a Draft PR because current code fails to replicate results in original S4 codebase. Running apples-to-apples (or as close as possible) results in ~66% accuracy (vs. 89% reported).

Possible reasons:

• Lack of custom LR per S4 Parameters (see #16 -- this function).
• Lack of LR Schedule (LR decay on validation/test accuracy plateau).

Resolves #8

[siddk]

Merge CIFAR --> just folded in the Optimization commits.

[siddk]

Updated results, run command is here.

Resolves #8

[siddk]

@srush - right now the replication command is just in the README; do we want to get rid of the unnecessary (other hyperparameter) runs?

[srush]

Yes, think we should clean up the readme to the working commands.

[siddk]

Done