Reference implementation for GGMLQuantizationType

[ngxson]

Currently, we already have all GGMLQuantizationType defined in quant-descriptions.ts

It may (or may not?) be interesting to have reference implementation to convert quantized types from/to F32.

This idea is actually brought to me from ggerganov/llama.cpp#6715 , which I was quite hesitate since there is no access to quantized type from gguf-py. Things changed a bit recently, since basic support for q8_0 has just been added to the convert.py script: ggerganov/llama.cpp#7234

The implementation will be easy for qX_0 types, a bit harder for qX_K and very hard for IQ series (with imatrix), so if we're considering doing this, that will be quite a long process. For this reason, I open this discussion to gather some ideas and feedback before actually working on this.