fix: use H200 default chunked-prefill-size of 8192

4096 was below SGLang's auto-detected default for H200 GPUs (<160GB),
which unnecessarily limited prefill throughput.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: PierreLeGuen

Qwen3.5-122B.yaml

@@ -54,7 +54,7 @@ x-sglang-qwen35-122b-common: &sglang-qwen35-122b-common
     --mem-fraction-static 0.88
     --context-length 262144
     --kv-cache-dtype fp8_e4m3
-    --chunked-prefill-size 4096
+    --chunked-prefill-size 8192
     --attention-backend flashinfer
     --schedule-conservativeness 0.5
     --reasoning-parser qwen3