Use opmath for sinc on CPU (#120311)
This aligns the implementation with CUDA and `torch.compile`
Fixes https://github.com/pytorch/pytorch/issues/118176 https://github.com/pytorch/pytorch/issues/49133
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120311
Approved by: https://github.com/jgong5, https://github.com/Chillee