fix: LL proto ar dispatch

[TennyWang1223]

Motivation

DeepSeek decode uses allreduce shapes (16, 7168) and (32, 7168) bf16 (224/448 KB). These fell through to the barrier-based ar_gfx1250_naive_unroll4
because the LL threshold was only 128 KiB. Raise it so these cases route to the faster flag-in-data ar_ll_gfx1250 kernel.

Technical Details

Changes in csrc/include/custom_all_reduce_gfx1250.cuh:

• kLLArMaxBytes: 128 KiB → 4 MiB
• kLLScratchCapBytes: 256 KiB → 4 MiB (scratch capacity to match)
• allreduce_ll block size: bytes < 65536 → 512 threads, otherwise 256
• ar_ll_gfx1250 __launch_bounds__: (256, 2) → (512, 1)

Test Plan

python op_tests/multigpu_tests/gfx1250_poc/test_gfx1250_allreduce.py
-s 16,7168 -s 32,7168 -d bf16 -g true -t 4 --check-arch

Test Result

tp_size	shape	dtype	size_kb	min_us	max_us	err
4	(16, 7168)	torch.bfloat16	224	8.884	9.172	0.000183
4	(32, 7168)	torch.bfloat16	448	10.066	10.415	0.000166

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
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 4753 --add-label <label>