test: reduce Kimi K3 B300 MegaMoE memory pressure

[alphabetc1]

Summary

• lower the Kimi-K3 B300 MegaMoE test static memory fraction from 0.85 to 0.82
• explicitly cap DSPARK max running requests and decode CUDA graph batch size to the CI GSM8K load
• keep the existing GSM8K accuracy gate while leaving headroom for FlashInfer autotune and CUDA graph capture

Context

PR #41092 hit CUDA OOM during FlashInfer MLA autotune in TestKimiK3B300MegaMoE.setUpClass, then NCCL _REDUCE_SCATTER_BASE timed out and rank 6 aborted:
https://github.com/sgl-project/sglang/actions/runs/36016352823/job/107760312725?pr=41092

This PR isolates the CI memory-headroom adjustment on top of main so we can validate it independently.

Test

• python3 -m py_compile test/registered/e2e/models/test_kimi_k3_b300.py
• git diff --check upstream/main..HEAD

---

CI States

Latest PR Test (Base): ⏳ Run #36105970820
Latest PR Test (Extra): 🚫 Run #36106026251
Latest PR Test (AMD ROCm 10): ⏳ Run #36105970745

[mmangkad]

Fixed by #41138