[programming][machine_learning] llama.cpp performance tuning

For unified memory systems (ex., AMD Strix Halo, Macs, etc.), use these options to lower RAM usage at the cost of performance:

```
  --cache-ram 0 \
  --ctx-checkpoints 1 \
```

or increase performance at the cost of RAM:

```
  --cache-ram 1 \
  --ctx-checkpoints 0 \
```

https://www.reddit.com/r/LocalLLaMA/comments/1sq94qx/comment/oh7qlju/