Add support for `rsqrt()` and `rcp()`

[fwyzard]

On CUDA there are two versions of rsqrt()

• rsqrt()
• __frsqrt_rn()

rsqrt() is accurate to 2 ULPs, while __frsqrt_rn() should be correctly rounded.

[VinInn]

ditto for "rcp" and for the double precision versions
(plus several other functions of a lesser use)

[VinInn]

Of course HIP/Rocm supports the "non intrisic" versions
( see https://rocm.docs.amd.com/projects/HIP/en/docs-5.7.0/reference/math_api.html )