Improve AMD accelerator example

[peterschmidt85]

Summary

• Rework the AMD accelerator example around fleets, inference, training, dev environments, Docker image, and metrics
• Use MI300X examples that can request at least four GPUs for inference and training
• Remove stale AMD links from the TRL and Axolotl training examples

[Bihan]

I recommend adding

volumes:
  - /checkpoints:/checkpoints

and setting --output_dir /checkpoints