New iteration on convert HF, fix some models, support Qwen3/Qwen3MoE

[francoishernandez]

We'll probably never have a perfect solution to handle every HF cases, but it doesn't hurt to keep rationalizing a few things.

Addressed topics

• centralize mappings and configs in separate file
• clarify encoder/decoder keys mapping (previously, decoder stuff would be in the root mapping, whereas encoder would be in the specific key)
• first shards params are transparently grabbed from mapping root, instead of relying on a fixed set which is a hassle to maintain
• move specific config flags to "config" key of main mapping
• simplify shards building loop (ongoing -- shall we loop on params/map instead of checkpoints?)

Some notes:

1. while testing this, I checked Mixtral quickly, and it appears to have been broken for a while (even before previous refactoring); not sure if we'll fix this here or later EDIT: MoE (Mixtral/Qwen3) seems fine after a few patches but AWQ is not -- though deprecated so not sure we want to dive back into it (might be better off investigating llm-compressor which replaces it)
2. Did not test all architectures yet (e.g. gpt2/nllb/xlmroberta) EDIT: only XLM-roberta not fully tested
3. transformer decoder refactoring a while ago introduced post_attention_layernorm, which should probably be made optional (e.g. phi-2) EDIT: introduced post_attention_layernorm flag (default True)