Skeleton for GGUF conversion (#2018)

Summary:
Starting a skeleton implementation

- Only llama for now. Will add new architecture inside gguf_util/converters/
- Only fp32. Will figure out the quantization.
- Reusing the existing llama code in examples to reduce duplication. For other architectures, there won't be much duplication.

- Currently converting to PyTorch, and then going through export, to_edge, to_executorch. But that's an implementation detail.

Pull Request resolved: https://github.com/pytorch/executorch/pull/2018

Test Plan: `python extension/gguf_util/convert_main.py --gguf_file="/Users/mnachin/models_gguf/OpenHermes-2.5-Mistral-7B-fp16.gguf"`

Reviewed By: shoumikhin

Differential Revision: D53982833

Pulled By: mergennachin

fbshipit-source-id: 5402c0de3e729e434763a5d6a390448603e77429