Fix a bunch of opinfo tests by using reference_in_float (#109089)
I set reference_in_float to be always True, ran the full opinfo test
suite, and observed which tests were now unexpectedly passing. However,
I didn't turn on reference_in_float by default in this diff because it
also creates some new failures.
Related: https://github.com/pytorch/pytorch/issues/105534
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109089
Approved by: https://github.com/eellison
ghstack dependencies: #109081