Fix optimized log_softmax along non-contiguous dim (#8666)

Fix log_softmax along non-contiguous dim

Pull Request resolved: https://github.com/pytorch/executorch/pull/8595

#8382 certainly didn't fix this problem (and added it on x86), but I don't think it was correct on ARM prior to that either.

Added a regression test.
ghstack-source-id: 268149462
@exported-using-ghexport

Differential Revision: [D69928884](https://our.internmc.facebook.com/intern/diff/D69928884/)

Co-authored-by: Scott Wolchok <swolchok@meta.com>