Add technical guide

[luciaquirke]

Conceptual guides

• add an example script for a data parallel multi-node slurm job
• explain how bias terms may be incorporated with design reasoning
• add guide on choosing between building an index and on-the-fly queries
• the preconditioner mixing feature and any other heuristic procedures

Performance

• which direction to tune each knob to increase speed and reduce accuracy
  • projection dim vs. none
  • preconditioning
  • dtype
  • faiss index query (https://arxiv.org/pdf/2012.15781 studied FAISS KNN to speed up IFs, we can cite)

Concepts from the broader literature

• explain self-influence and the best way to calculate it
• explain the best way to replicate the Cohere paper
• explain the best way to replicate TrackStar
• explain the best way to replicate LESS

[luciaquirke]

• adafactor weight matrix optimization (see above)
• skip_index doesn't reduce VRAM but it doesn't increase the run speed due to less IO
• precision reduces disk IO but preconditioners stay in float32

[luciaquirke]

https://github.com/EleutherAI/bergson/pull/93/changes
Configs are creeping into the core code, not sure if we should change this