Update comment on nvc++ workaround in mem_fence()

[nmnobre]

Hi @illuhad,

It seems Nvidia has indeed fixed the compiler errors with empty if target statements.
See #767 (happy to close if you are) and #798.

Cheers :)

[nmnobre]

On second thought, I think this is overcomplicating things and adds nothing. Besides, it's the wrong place for a fix like this...
I would suggest either:

1. Adding a note to the comment saying it only applies to versions <= 22.7;
2. Editing cuda_backend.hpp to add printf(""); or int dummy; (void)dummy; and moving the compiler bug comment there;
3. Amending the documentation to say we require version 22.9 or newer.

[illuhad]

I think it would suffice to just add to the comment that this only affects these older nvc++ versions. We don't need to put in a lot of effort just to be able to express a no-op in an if target path :-)

[illuhad]

Thanks!