test: DSv4-Pro B300 vLLM - widen conc to [1,2,4,8,16,32,48,64,70]

Full power-of-2 sweep up to 32, then 48, 64, 70 to walk through the
likely cliff region with finer steps. Same 4 parallelism configs,
still no offload, v0.20.2 image. 36 jobs total.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: cquil11

.github/configs/nvidia-master.yaml

@@ -2844,10 +2844,10 @@ dsv4-fp4-b300-vllm:
     agentic-coding:
     - duration: 1800
       search-space:
-      - { tp: 4, offloading: none, conc-list: [16, 32, 64, 70, 76] }
-      - { tp: 8, offloading: none, conc-list: [16, 32, 64, 70, 76] }
-      - { tp: 4, ep: 4, dp-attn: true, offloading: none, conc-list: [16, 32, 64, 70, 76] }
-      - { tp: 8, ep: 8, dp-attn: true, offloading: none, conc-list: [16, 32, 64, 70, 76] }
+      - { tp: 4, offloading: none, conc-list: [1, 2, 4, 8, 16, 32, 48, 64, 70] }
+      - { tp: 8, offloading: none, conc-list: [1, 2, 4, 8, 16, 32, 48, 64, 70] }
+      - { tp: 4, ep: 4, dp-attn: true, offloading: none, conc-list: [1, 2, 4, 8, 16, 32, 48, 64, 70] }
+      - { tp: 8, ep: 8, dp-attn: true, offloading: none, conc-list: [1, 2, 4, 8, 16, 32, 48, 64, 70] }
 
 dsv4-fp4-b300-trt:
   image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-9aa3715