[CI] Add inference run for the performance dashboard (#98174)
Summary: Remove fp32 training performance run and trade for amp inference
performance run.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98174
Approved by: https://github.com/huydhn