Affine Lock 5.0 Port and Cleanup and CUDA Version

[Kinvert]

This isn't probably fully done. But this is a lot of the cleanup before you might have agents keep tweaking things. It's sweeping now. Faster than the 4.0 version.

[Kinvert]

A small change to core fixing a potential bug, feel free to ditch that part of it.

[Kinvert]

Potential improvements:

• Log depth_D_attempts, it's already calculated and can potentially indicate what depths might be memorized
• Experimenting with step_grace (step_grace == 0 requires perfect optimal solve) scaffolding already exists
• Can potentially expand the horizontal size, like the number of bits in the affine lock (maybe a big change though)
• Add curriculum modes in the ini. I think this game can have multiple versions. Right now it has to solve sequentially or start over. Could add where it is only demoted 1 curriculum level on a failure.
• Put extra time in the generator and docs to make sure it's effective and easy to use
• Add hash gen to generator so we can know we're training on the same .bin
• Build in test train isolation, keep in mind D16 seems to have only 100,548 perfect solutions mathematically
• Formulate a perf value that keeps its meaning over different changes like curriculum levels, curriculum demotion, etc. Good perf must decide if d8 0.25 is better than d6 0.80 for example, and stay apples to apples when we change curriculum handling.
• Allow more reward schemes
• Adaptive but deterministic curriculum rather than current fixed curriculum, Teacher-Student, Prioritized Level Replay
• Potential PufferLib Core - When training on .bin we might make Test/Train and have eval pull from Test.