handle tokenization more correctly, clean up CI

[charlesfrye]

No description provided.

[charlesfrye]

This PR improves the CI job (and SQL QA training in general) in several ways. Tokenization is the most important. Tokenization is still not guaranteed to be identical between axolotl and vllm, which merits some more investigation.

Summary of changes:

• All of the SQL QA configs now have custom tokens (e.g. [SQL]) added to improve the tokenization of inputs.
• All models are switched over to the default, AutoTokenizer, which defers lookup of the tokenizer to hugging face. This substantially improved the quality of results, but I didn't root cause.
• The prep for the "memorization test" in CI is adjusted to speed up runs.
• llama-2.yml is removed, since we now have llama-3.yml.
• Extra config keys are removed from some of the configs.
• Switches to A100s instead of H100s as the default GPU to reduce time spent queuing, in exchange for slightly slower training.