test(ark): add base+deq(ms) column to MoE perf benchmarks
Time the per-step weight dequantization cost separately and report
it alongside the matmul-only baseline as base+deq(ms). Use
(base+deq)/ark as the speedup denominator so the comparison reflects
the realistic cost a pipeline keeping weights in quantized storage
would pay if it reused a stock matmul baseline. FP rows have
deq_ms == 0 so base+deq == baseline.
Updates both test_moe_prefill_perf.py and test_moe_decode_perf.py
plus README_MOE_PREFILL_PERF.md.
