Implement test to quantify the impact of different features on gradient approximation

**Is your feature request related to a problem? Please describe.**

There is currently no test to quantify the error in gradients introduced by different approximations, e.g.:

* Quantization of activation checkpoints
* Cell-based computation [any?]
* Autograd saved tensors packing [any?]

**Describe the solution you'd like**

We need a sensible metric (or several) for gradient errors; a setup where gradients can be computed exactly; a way to get to realistic weights.

No files, tests, or entry points are identified. First locate the implementations and tests for activation checkpoint quantization, cell-based computation, and autograd saved-tensor packing, then define a gradient-error metric, an exact-gradient setup, and realistic weights; done means the test quantifies each approximation's impact.

Written by the indexing model from the issue text.