Support zero-point decompression for asymmetric quantization (packed compressor)

[Etelis]

When using asymmetric quantization with PackedQuantizationCompressor, zero-points are packed during compression but were not unpacked during decompression. This prevented loading and inference for models using GROUP or CHANNEL strategies.

This tracks the feature request in vllm-project/llm-compressor#1704 and proposes adding zero-point unpack support in decompress_weight.

References:

• [Feature] Support Zero-point Decompression llm-compressor#1704 ([Feature] Support Zero-point Decompression llm-compressor#1704)