[Feature] Compact IntDim-L/G experts and add a ragged Qwen vLLM GPU backend #29

Problem

IntDim-L/G can keep different neuron counts in each expert, but the current exporter only supports uniform-width structural pruning. Zero masking retains the original dense matrix shapes, so stock vLLM does not skip those computations.

Scope

Implement a first GPU-only Qwen MoE path on the pinned unified environment (PyTorch 2.13.0+cu130, Transformers 5.17.0, vLLM 0.29.0):

• Versioned checkpoint metadata describing each layer's per-expert widths and compact weight layout.
• Physical compaction using the existing calibration-based IntDim-L/G selection, with an explicit compatible HF reload path.
• An opt-in vLLM model integration and a Triton GPU expert kernel that reads the actual expert widths; preserve routing, attention and shared experts.
• Start with BF16, single GPU / TP=1; reject unsupported configurations explicitly.

Acceptance

• GPU tests compare compact expert outputs, full-model logits and reloaded weights against the corresponding zero-mask reference.
• Cover both layer and global selection, unequal/unaligned widths and invalid metadata.
• Run vLLM inference on B200 using a small Qwen MoE fixture, then a complete pretrained Qwen MoE checkpoint if resources permit. Report fixture and pretrained evidence separately, including exact environment, settings and limitations.
• Record reproducible commands and committed test code. Successful inference does not by itself establish a throughput speedup.

Related: #25, #9. Existing uniform-width IntDim-E and stock runtime behavior must remain available.