Add GLM GQA FP8 KV paged attention test

[ThomasNing]

Summary

• Add a GLM-style paged_attention_ragged regression test for GQA FP8 KV cache
• Covers BF16 query/output, FP8 K/V cache, NHD layout, page size 1, 96 query heads, 8 KV heads, and head size 128
• Compares native FP8-KV decode against the same FP8 values dequantized to BF16

Why

GLM-4.5 style serving uses GQA with 96 Q heads / 8 KV heads / D=128. This pins official AITER coverage for direct FP8 KV decode so framework integrations can route to AITER without dequantizing the cache first.

Test

• On MI300X gfx942 with lmsysorg/sglang:v0.5.9-rocm700-mi30x:
  python3 -m pytest -q op_tests/test_pa_ragged.py::test_paged_attention_ragged_glm_gqa_fp8_kv_nhd -s
• Result: 1 passed, 2 warnings in 25.67s

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 3609 --add-label <label>

[ThomasNing]

I should add up one more SGLang patch in the following to prevent it be dequantized.