[Microbench] loss.soft_margin_loss forward/backward have performance regression

[huaiyuzh]

Zhong, Ruijie found this regression. File an issue to track this perf regression on torch-xpu-ops.

[Summary]

According to 25ww07 weekly test, we check the torch2.6 found some ops have perf regression compared with 25ww06 weekly results

[op list]

loss.soft_margin_loss
loss.soft_margin_loss_backward

[Results]

https://intel-my.sharepoint.com/:x:/r/personal/ruijie_zhong_intel_com/_layouts/15/Doc.aspx?sourcedoc=%7BD7790E78-B70B-404F-A510-4A1A27973E41%7D&file=ww35_microbench_report.xlsx&action=default&mobileredirect=true