Implement AtomicFAddEXT for the CUDA BE

[AGindinson]

After 4fdbfae, there are preparations to switch atomic fetch_add/fetch_sub FP implementations to using the new SPIR-V operand. Providing a "native" implementation in the CUDA BE would enable us to use the leveraged function for NVPTX targets as well (#if !defined(__NVPTX__) macros would have to be removed to achieve this).

[ldrumm]

I think this is now implemented. It looks like @AGindinson did the meat of this work in 37a9a2a

Additionally, relevant libclc support went in in the following PRs:
#4820
#4853
#5025
#5191

@AGindinson is there anything missing? Perhaps we can close this?

[AGindinson]

@AlexeySachkov, could you please help with evaluating this one?

[npmiller]

@AlexeySachkov @AGindinson any updates on this?

[AlexeySachkov]

@AlexeySachkov @AGindinson any updates on this?

Not really. Both of us are not directly working on CUDA, so this item is a lower priority for us both. Feel free to pick it up. I'm also fine with closing it if we believe that everything is implemented already

[npmiller]

From a quick look into the headers I don't see any #if !defined(__NVPTX__) usage for fetch_add/fetch_sub, so I believe this is implemented, I'm fine with closing it, what do you think @ldrumm ?