Fix CUDA issue with irrational constant

[devmotion]

I think upstream fixes for #73 might take a while. Since we only use a heuristic here, it should be fine to just promote the irrational to the same (floating-point) type of the input x before comparing it with x.

@Red-Portal can you confirm that this PR fixes your issue? Seemed to be the case for me at least.

[Red-Portal]

Yes, can confirm!

julia> @eval LogExpFunctions begin
           function log1mexp(x::Real)
           # Use explicit `oftype(..)` instead of just `loghalf` to avoid CUDA issues:
           # https://github.com/JuliaStats/LogExpFunctions.jl/issues/73
           if x < oftype(float(x), IrrationalConstants.loghalf)
               return log1p(-exp(x))
           else
               return log(-expm1(x))
           end
       end
       end
log1mexp (generic function with 1 method)

julia> log1mexp.(CUDA.randn(3))
3-element CuArray{Float32, 1, CUDA.Mem.DeviceBuffer}:
  -0.4460078
  -0.6540878
 NaN