Small operatorbench improvements (#103110)
- Don't copy inputs in cudagraphs wrapping, since the copies will distorts timing and triton do_bench will clear cache anyway
- Don't skip op if there is a fallback, since we have both fallbacks and lowerings for some ops
- Add option for channels last
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103110
Approved by: https://github.com/desertfire
