WIP Distribution Visualisation to help with FP8 work & beyond

[ademeure]

Not ready for integration at all / still very hacky, bunch of unsolved issues I am not sure where code should go etc.:

• need to find a way to make it pollute the code less with all of those generate_analysis() calls everywhere
• the declarations should obviously not be at the top of train_gpt2cu, maybe a separate file?
• the rest of the code probably shouldn't be in cuda_utils.h, maybe together in that new file?
• make it work with multiple steps (stopping train_gpt2cu after one micro-step now, so it can't get any worse!)
• need to clean up the mess that is the HTML/JS file created through 100 iterations of 4o/Opus...

... but...

It's pretty cool! :)