[llvm] [NVPTX] Fixup some issues introduced by 128-bit atomics (PR #155921)

[Artem Belevich via llvm-commits]

Artem-B wrote:

It appears that after this patch we also see a new ptxas test failure in llvm/test/CodeGen/NVPTX/atomics-b128.ll
Looks like ptxas is unhappy about `.seq_cst` modifier we're generating now.

```
[  5] ; RUN: %if ptxas-12.4 %{ llc < %s -mcpu=sm_90 -mattr=+ptx84 | %ptxas-verify -arch=sm_90 %} [FAIL]
 llc < /build/work/67189cfec901062d332d75ef1db0b2555386/google3/runfiles/google3/third_party/llvm/llvm-project/llvm/test/CodeGen/NVPTX/atomics-b128.ll -mcpu=sm_90 -mattr=+ptx84 | /build/work/67189cfec901062d332d75ef1db0b2555386/google3/runfiles/google3/third_party/gpus/cuda/_virtual_includes/_stage_runtime/third_party/gpus/cuda/bin/ptxas -arch=sm_60 -c -o /dev/null - -arch=sm_90 
ptxas warning : incompatible redefinition for option 'gpu-name', the last value of this option was used
ptxas /tmp/tmpxft_00001fb0_00000000-0_stdin, line 1082; error   : Unknown modifier '.seq_cst'
ptxas fatal   : Ptx assembly aborted due to errors
Command failed: exit status 255
```

https://github.com/llvm/llvm-project/pull/155921