No need for atomicAdd for float2, conflicts with CUDA 12.1

[nshmyrev]

Since 12.1 CUDA added float2 version of atomicAdd

https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#atomicadd

which causes build error:

#7 392.3 feature-online-batched-cmvn-cuda-kernels.cu(34): error: cannot overload functions distinguished by return type alone
#7 392.3   __attribute__((device)) inline void atomicAdd(float2 *addr, float2 val) {

as far as I see currently this function is not used, we can remove it safely

[nshmyrev]

Similar old problem

https://github.com/kaldi-asr/kaldi/issues/1214

[jtrmal]

LGTM, if the function is not used, we don't have to worry about conditional compilation

[jtrmal]

Thanks!