Reland support premul sum for xccl

[Chao1Han]

Reland #1948

disable_e2e
disable_ut

[Copilot]

Pull request overview

This PR reintroduces XCCL support for ReduceOp::PREMUL_SUM when building against oneCCL >= 2021.17 by adding version gating, reduction-op construction helpers, and RAII cleanup for custom reduction handles.

Changes:

• Add compile-time ENABLE_XCCL_PREMUL_SUM_SUPPORT based on oneCCL version macros.
• Introduce RAII wrappers and unpack helpers to create/destroy PREMUL_SUM reduction ops for both CCL “V1” and oneCCL “V2” APIs.
• Update XCCL collective call sites to pass datatype/communicator needed to build PREMUL_SUM ops.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
src/xccl/xccl.h	Adds PREMUL_SUM enablement, RAII reduction wrapper, and PREMUL_SUM mapping logic in getXcclReduceOpV1/V2.
src/xccl/xccl.cpp	Updates allreduce/reduce/reduce_scatter wrappers to pass datatype + communicator into the reduce-op selection helpers.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Performance outliers, please check!

• 🔴 [-1, 80%), should be regression

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
torchbench_bfloat16_training	Background_Matting	0.744062	0.682433
torchbench_bfloat16_training	pytorch_unet	0.765713	0.725342
torchbench_bfloat16_training	alexnet	0.772230	0.754138
torchbench_bfloat16_training	resnet50	0.755595	0.764938
torchbench_bfloat16_training	nvidia_deeprecommender	0.750006	0.770794
torchbench_bfloat16_training	shufflenet_v2_x1_0	0.783204	0.858808
torchbench_bfloat16_training	LearningToPaint	0.686767	0.862129
torchbench_bfloat16_training	vgg16	0.752418	0.864969

• 🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
torchbench_bfloat16_training	mobilenet_v2	0.843263	0.828889
torchbench_bfloat16_training	resnet18	0.850908	0.839607
torchbench_bfloat16_training	BERT_pytorch	1.011453	0.895164
torchbench_bfloat16_training	mnasnet1_0	0.879395	0.946367
torchbench_bfloat16_training	squeezenet1_1	0.821079	0.946886
torchbench_bfloat16_training	resnext50_32x4d	0.895986	1.002764