Use an extra bit of entropy in the Float64 uniform 0-1 distribution.

[nhz2]

When looking over #707 I think the following is a slight improvement for Float64 u01.
This PR subtracts prevfloat(1.0) to avoid returning zero from u01 while still using 52 bits.

cc @maleadt

[nhz2]

I think the Float64(::UInt64) being expensive was a red herring. The actual issue was the Float64(2)^(-64) not being constant propagated.

@inline function u01(::Type{Float64}, u::UInt64)::Float64
    fma(Float64(u), Float64(2)^Int32(-64), Float64(2)^Int32(-65))
end

Seems to be fine performance wise. Ref: https://github.com/medyan-dev/PhiloxRNG.jl/tree/v1.1.1#gpu--nvidia-geforce-rtx-3080-nsvalue-n--100000000

[maleadt]

The actual issue was the Float64(2)^(-64) not being constant propagated.

This was fixed in JuliaGPU/CUDA.jl#3098, which I encountered as part of porting your RNG.