Add `fast` versions for `sqrt`, `rsqrt`, and `recip`

[stijnh]

This is a list of the fast approximate functions that CUDA supports that we should also support:

f32:

• sqrt
• rsqrt
• rcp
• sin
• cos
• div
• log
• log2
• log10
• exp
• exp2
• exp10

f64

• rcp
• rsqrt

[stijnh]

See 4d3d356