Replace exp(x) - 1 with expm1(x) (#92154)
This offers improved precision near zero where `exp(x)` is `1 + O(x)` and doing
`(1 + O(x)) - 1` will truncate anything below the float epsilon to zero.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92154
Approved by: https://github.com/lezcano
