Break our dependencies into a common core and optional subgroups

[nenb]

What problem are you trying to solve?

• Make it easy for full-stack developers, small organisations and MongoDB users in general to deploy SuperDuperDB as part of their app. Currently, a minimum installation of SuperDuperDB requires >5GB of disk space, and a surprisingly large amount of RAM - certainly greater than the free tier limits that platforms like fly.io offer for deploying apps (@thejumpman2323 has linked this RAM demand to torch from his past experience).

Why should we solve it?

• We want to increase adoption of SuperDuperDB - we want to make it as easy as possible for users to deploy SuperDuperDB as part of their app.

How do you propose to solve it?

• Make torch an optional dependency as this is likely 80% of the work. Most of the work has already been done by Rename and reorganize code (fix #540) #578. All that remains is to remove a small number of imports to torch in the core code, and then to update the pyproject.toml to include a new section for optional dependencies.
• Extend the torch work to the remaining subgroups (these need TBD)
• Extend CI so that we install all optional subgroups independently and check that the imports are still valid

What could go wrong?

• We realise that torch is essential for our core functionality in some unexpected way
• We silently break installation of our package (can be prevented once the CI is ready)