Use at::Vectorized in optimized log_softmax

Pull Request resolved: https://github.com/pytorch/executorch/pull/8382

This should allow us to enable this op in OSS, because Vectorized handles any Sleef issues for us as needed. (I considered going straight to sharing the PyTorch core implementation, but we need parallel_for enabled for that and this improvement is easy enough to make.)

Differential Revision: [D69473208](https://our.internmc.facebook.com/intern/diff/D69473208/)
ghstack-source-id: 267044107

Co-authored-by: Github Executorch <github_executorch@arm.com>