[optim][muon] fix momentum buffer aliasing in Newton-Schulz (#194664)

`_zeropower_via_newtonschulz` happens in-place.
* before: `grad.bfloat16()` returns itself when already bfloat16, and then normalized it in place
via `div_`
* after: use `grad.to(dtype=torch.bfloat16, copy=True)` so the input is never aliased

Pull Request resolved: https://github.com/pytorch/pytorch/pull/194664
Approved by: https://github.com/janeyx99