[CD] Set different time measure for microbench

[RUIJIEZHONG66166]

Enhance the op microbench test with device, test method and iter number ondemanded control.
disable_ut

[RUIJIEZHONG66166]

Ondemand test results: https://github.com/intel/torch-xpu-ops/actions/runs/16987046730
If Profile Change = "↓" or "↑", it means the regression or improvement exceed the threshold; if no Profile Change, it means the regression or improvement do not exceed the threshold.
Some outlier like case in scan.cumsum have regression because it compared with main branch results, and main branch's test scripts "scan.cumsum.py" has type issues which fixed by this PR. Will refresh the baseline after this pr merged.

[RUIJIEZHONG66166]

After rebasing, ondemand test results: https://github.com/intel/torch-xpu-ops/actions/runs/17570650066

[chuanqi129]

After rebasing, ondemand test results: https://github.com/intel/torch-xpu-ops/actions/runs/17570650066

Hi @RUIJIEZHONG66166 , I have checked the on-demand test summary, there are a lot of regression/improvement cases, does it cause by test variance? Please hide improvement and mixed changes tables by default. And if we run 2 methods, let's mark those regression both on 2 metrics as regression cases

[RUIJIEZHONG66166]

After rebasing, ondemand test results: https://github.com/intel/torch-xpu-ops/actions/runs/17570650066

Hi @RUIJIEZHONG66166 , I have checked the on-demand test summary, there are a lot of regression/improvement cases, does it cause by test variance? Please hide improvement and mixed changes tables by default. And if we run 2 methods, let's mark those regression both on 2 metrics as regression cases

Updated the comparison method:

1. Hided improvement and mixed changes tables by default.
2. Mark those regression both on 2 metrics as regression cases.

For these regression/improvement cases, the root cause may be the baseline commit number has not been updated due to an issue with write permissions in the workflow. I set the baseline to the latest on-demand test results(https://github.com/intel/torch-xpu-ops/actions/runs/17570650066) and did a new round of on-demand testing(https://github.com/intel/torch-xpu-ops/actions/runs/17669054947) compared with 17570650066, which showed no significant regression or improvement.