The CUDA target needs to correctly emit the convergent attribute

[gmarkall]

As described in the LLVM documentation on Convergent Operation Semantics.

Incorrectly failing to emit it is not a problem at present, but it may be a problem in the future.

Conservatively, we can be correct by emitting it on all functions and inline assembly - however there is probably more nuance required if we want to emit it only as-needed.