Fix wrong group usage in blocking p2p

[Chao1Han]

In the default blocking P2P mode, there is no need to wrap P2P operations with group APIs. In oneCCL 2021.17, this pattern can actually lead to a deadlock when a group contains only send or only recv. Removing the group calls around the P2P operations resolves this issue.
disable_e2e
disable_ut

[Copilot]

Pull request overview

This PR fixes a deadlock issue in blocking P2P operations by removing unnecessary group API wrapping. The change addresses a problem where oneCCL 2021.17 can deadlock when a group contains only send or only recv operations.

• Removed ccl::group_start() and ccl::group_end() calls around P2P operations in blocking mode
• Re-enabled previously skipped test_batch_isend_irecv test that was hanging due to this issue

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/xccl/ProcessGroupXCCL.cpp	Removed group API calls wrapping P2P operations to prevent deadlock
test/xpu/skip_list_dist.py	Re-enabled previously skipped test that was failing due to the deadlock issue

---

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

[pkourdis]

LGTM.

[github-actions]

Performance outliers, please check!

• 🔴 [-1, 80%), should be regression

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
torchbench_bfloat16_training	resnet18	0.887445	0.788249

• 🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
huggingface_float16_training	XLNetLMHeadModel	0.99703	0.899677

[github-actions]

Performance outliers, please check!

• 🔴 [-1, 80%), should be regression

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
torchbench_bfloat16_training	resnet18	0.908869	0.770172