[dist] reuse comm groups when possible for memory saving

[gbyu-amd]

Motivation

The comm group reuse was originally done in ROCm/ATOM#2050. Now implement it in aiter side to make it safer.

The purpose is to reuse comm buffers for different comm groups when they share the same ranks and collectives are issued in sequence. This can reduce considerable non-torch memory allocation, therefore leaving more capacity to KV cache.

Before:

[atom 21:34:21] Memory budget: total_gpu=287.98GB, free=63.91GB, utilization=0.86, budget=247.67GB, peak_torch=192.92GB, non_torch=30.21GB, cudagraph_est=0.40GB, safety=5.76GB, available_for_kv=18.37GB, block_bytes=1769472, num_kvcache_blocks=6832

By reusing the comm groups with this PR, non-torch memory now only takes 13GB out of the budget, almost doubles the kv cache:
After:

[atom 04:36:43] Memory budget: total_gpu=287.98GB, free=80.78GB, utilization=0.86, budget=247.67GB, peak_torch=192.91GB, non_torch=13.40GB, cudagraph_est=0.40GB, safety=5.76GB, available_for_kv=35.20GB, block_bytes=2138112, num_kvcache_blocks=13727

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[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:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
multigpu	Aiter multi-GPU tests on the 8-GPU runner
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 5111 --add-label <label>

PR title tags:
Component tags ([Triton/Gluon], [HIP], [CK], [ASM], ...) are added to the PR title automatically from the changed files and re-synced on every push — change-type tags like [fix]/[Perf] and op tags like [MLA] are left untouched. Add the no-auto-title label to opt this PR out of title tagging.