[#17723][feat] Parse multiple compressed-tensors config groups

[mihai-chiorean]

Summary

Parse every compressed-tensors config_groups entry into a MIXED_PRECISION global config plus an explicit per-module quant_config_dict. Preserve target ordering, reject unmatched targets, and keep each module configuration independent. Resolve module names from local safetensors files or a Hugging Face Hub index.

This is required by unsloth/Qwen3.8-27B-NVFP4, which mixes packed NVFP4 dense MLPs with rowwise FP8 attention, GDN, later MLP, and lm_head modules.

Related to #17723.

Details

• Preserve homogeneous single-group behavior.
• Parse nvfp4-pack-quantized as native W4A4 NVFP4; activation scales are part of that format and must not be discarded by relabeling it W4A16.
• Apply producer ignore entries while materializing the authoritative multi-group map without copying them into recursive global exclusions, which could shadow explicitly quantized children.
• Preserve only modules_to_not_convert as a recursive exclusion in the multi-group path.
• Read sharded indexes and single safetensors headers from local paths or Hub IDs; fail before model construction if the per-module map cannot be resolved.

Testing

• pre-commit run --files on all changed files
• Local-index, Hub-index, and ModelConfig.load_hf_quant_config wiring tests
• Parser and model-config tests included in a 200-pass integrated Spark run
• Full Qwen3.8-27B NVFP4 SM121 smoke reached native trtllm::nvfp4_gemm and generated coherent output

Dependencies

This parser is generic and can merge independently. Qwen3.8 end-to-end loading additionally needs the packed-name, dense-MLP path, and FP8 lm_head PRs linked from #17723.

Interaction with #17551

#17551 merges producer ignore into recursive exclusions for homogeneous configs. If it lands first, this PR must retain an explicit MIXED_PRECISION carve-out: the per-module map is authoritative for multi-group configs, and recursively re-injecting ignore can shadow a quantized child.