Disable vml's abs and log1p (#92113)
I noticed that `torch.log1p` is ridiculously slow compared to `torch.log`
on CPU, and looking at the assembly it seems vsLog1p doesn't use any
vector instructions. I saw the same for abs, though AFAICT this is
dead code anyway as `abs` is implemented with `cpu_kernel_vec`.
Locally I see a 14x speedup in `torch.log1p`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92113
Approved by: https://github.com/jgong5
