Improve Conversion Utilities

[haileyschoelkopf]

[Draft for now]

This PR seeks to improve the state of conversion / checkpoint import+export tools in GPT-NeoX.

Intended changes are:
NeoX to HF:

• Unify the previously-separate PipelineModule and Sequential conversion scripts
• Address some NeoX-to-HF QoL features (manually override checkpoint's save precision, allow to skip tokenizer upload (if one is using Tiktoken, say)
• Allow for Mistral + Llama models, and GQA in these

Meta Llama / Mistral to NeoX:

• enable + test GQA
• Add Llama 2 loading instructions

Tests:

• convert_module_to_hf.py and convert_neox_to_hf.py output precisely the same files, even under tensor parallelism (NeoX)
• convert_sequential_to_hf.py and convert_neox_to_hf.py output precisely the same files, even under tensor parallelism (NeoX)
• Llama model conversion functions, in the absence of GQA
• Llama model conversion functions, with GQA

[Quentin-Anthony]

All comments addressed over DM. Great work!