[NVPTX] Add support for atomic add for bf16 type

[akuegel]

atom.add.noftz.bf16 is supported since SM 9.0 and PTX 7.8

[Artem-B]

LGTM.

Do we need to worry about denormals? I.e. we only have noftz variant of the atomic, which potentially will give us different results that a regular addition would, depending on denormals handling mode. Do we need to add an extra op that folds denormals to zero when normal FP ops would be expected to do so?

I'm way out of my depth on FP handling nuances here. We need to ask an FP expert.

[akuegel]

@Artem-B Do you know someone we could ask who is a floating point expert?
Note that in my previous change where I added support for f16 type, I also didn't handle the case with ftz mode. So do I need to revert that previous patch?

[Artem-B]

Do you know someone we could ask who is a floating point expert?

@majnemer Do you have an opinion on atomics and denormals handling for bf16/f16? Who would be the right folks to consult on that?

[majnemer]

I think it is OK for us not to do anything special for noftz.

It is a bit tricky to fix it post-hoc because of cases like:

  x = 0
  x += 0x1.02p-126
  x += -0x1p-126

Here, none of the inputs were subnormal but x ends up as 0x1p-133 which is subnormal.

Also, LLVM semantics allow us to return subnormal results instead of flushing.

LGTM.