[Experiment] Table 5 (optional): iso-active-parameter comparison #19

Answers Q2: does Qwen3.5-122B-A10B pruned at p = 50% (best IntDim variant) beat the dense Qwen3.5-9B with a similar active-parameter budget?

Metrics: total and active parameters, disk, throughput, GPQA-Diamond, IFEval, IFBench, MMLU-Pro, C-Eval.
Qwen3.5-9B is evaluated as a reference only and is not pruned.

Depends on: #9, #15

Done when

• The three columns of Table 5 are filled.

Part of #2