[module_moe_asm] refactor and rm torch

[amd-ruitang3]

Summary

Remove the torch dependency from module_moe_asm's C/HIP sources.
The [module_moe_asm] build time drop from 33.8s to 16.4s (-51.5%)

Compile-time improvement (single-file hipcc, gfx950)

source file	before (torch)	after (torch-free)	saved
topk_softmax_kernels.cu	29.6 s	12.5 s	−17.1 s (58%)
topk_softmax_kernels_group.cu	27.1 s	9.8 s	−17.3 s (64%)
moe_fused_gate.cu	19.4 s	5.4 s	−14.0 s (72%)
moe_align_block_size_kernels.cu	19.3 s	1.8 s	−17.5 s (91%)
moe_op_pybind.cu	19.8 s	5.6 s	−14.2 s (72%)
total (serial)	115.2 s	35.1 s	−80.1 s (70%)

[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 4368 --add-label <label>