Add rsqrt as a native triton language function

[pawelszczerbuk]

Adding support for rsqrt in Triton. Also adding tests for unary math ops.

[ThomasRaoux]

Nice :)

[jlebar]

OOC what does this lower to in PTX?

https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#floating-point-instructions-rsqrt is probably what people want, but it's "approx".

[pawelszczerbuk]

OOC what does this lower to in PTX?

https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#floating-point-instructions-rsqrt is probably what people want, but it's "approx".

Yup, rsqrt.approx. We can add _rn, _rz versions of the function as we have done for other, if necessary.

[jlebar]

Should we add a warning to the documentation that this is approximate and 1/sqrt(x) may be more accurate? (Maybe the Triton function should even be called rsqrt_approx?)

[jlebar]

We can add _rn, _rz versions of the function as we have done for other, if necessary.

I'm not sure these exist for rsqrt.approx?