[llvm] [NVPTX] Support llvm.{exp2, log2} for f32/f16/bf16 and vectors (PR #120519)

[Artem Belevich via llvm-commits]

Artem-B wrote:

Note that I was specifically talking about `exp2`.

`log2` in CUDA gets forwarded to libdevice and does have an elaborate implementation there. It does *not* get lowered to `lg2.approx`. Approximate log2 should remain disabled by default.

https://github.com/llvm/llvm-project/pull/120519