Inference queue throughput

One GPU worker; alternate batch order each round. Queued latency includes waiting time. A throughput change is not a claim that FIFO accelerates model inference.

Run	Engine	Checkpoint	Clients	Rounds	Sequential s	Queued s	Throughput change %	Identical WAVs	Peak queue	Pass
20261025-050404-08222176	faster	0.6B-CustomVoice	4	5	77.421	79.195	-2.24	20/20	3	True
20261025-053632-9d5b8428	faster	1.7B-CustomVoice	4	5	88.723	88.070	0.74	20/20	3	True