Restore to previous state with copy-on-write

Apologies if this isn't the right place for this kind of feature request.

As a Jupyter user, I would find it _incredibly_ helpful if there was an "undo" for kernels. Frequently, I run something, only to realize it was the wrong thing and I want to go back one cell. If the bad cell had any destructive effect, this means I need to restart my kernel and run all the cells above _again_. If those cells took 3 hours to run, I need to wait 3 hours again. Even if I don't think the bad cell had any destructive effect, if I fix up the notebook mid-stream others (and myself) won't be 100% sure that the notebook would work with a simple "restart and run all."

I wish there were a command like "Kernel -> Save Checkpoint" and then the ability to restore back to checkpoints. In Jupyter, if this worked in tandem with the checkpointing of the notebook file itself, it could create a seamless experience: I create a checkpoint, make a mistake, restore both kernel and notebook to the checkpoint, fix the mistake and keep running, and I end up with _exactly_ the notebook I would have had if I hadn't made the mistake in the first place. (On the Jupyter side I'd also like an option that automatically checkpointed after each cell, and one that forks off into a console instead of checkpointing the notebook file.)

Technically, it seems like this could be achieved through `fork`'s copy-on-write mechanism. For many notebooks that don't have much volatile state that is rewritten multiple times, even checkpoints after every cell could have near-zero memory overhead -- they would just be forks of the kernel process at that point, using the same pages in memory as the active kernel process unless they are subsequently modified.

The issue does not name files, tests, or entry points. Start by investigating how the Jupyter kernel process could use fork's copy-on-write mechanism for checkpoints, including restoring kernel and notebook state; done would provide a reliable checkpoint and restore workflow without rerunning earlier cells.

Written by the indexing model from the issue text.