Use ternary operator instead of ifelse

[devmotion]

In benchmarks of JuliaStats/StatsBase.jl#714 and the examples shown below it seems that on my computer the ternary operator is sometimes faster than ifelse (which evaluates all branches) and never slower:

With the latest release:

julia> using LogExpFunctions, BenchmarkTools, Random

julia> Random.seed!(1234);

julia> out = Vector{Float64}(undef,100_000);

julia> x, y = rand(100_000), rand(100_000);

julia> @btime map!(xlogx, $out, $x);
  717.358 μs (0 allocations: 0 bytes)

julia> @btime map!(xlogy, $out, $x, $y);
  780.322 μs (0 allocations: 0 bytes)

julia> @btime map!(logistic, $out, $x);
  745.426 μs (0 allocations: 0 bytes)

julia> @btime map!(logaddexp, $out, $x, $y);
  2.502 ms (0 allocations: 0 bytes)

julia> @btime sum(xlogx, $x); # similar to https://github.com/JuliaStats/StatsBase.jl/pull/714
  780.203 μs (0 allocations: 0 bytes)

With this PR:

julia> using LogExpFunctions, BenchmarkTools, Random

julia> Random.seed!(1234);

julia> out = Vector{Float64}(undef,100_000);

julia> x, y = rand(100_000), rand(100_000);

julia> @btime map!(xlogx, $out, $x);
  703.294 μs (0 allocations: 0 bytes)

julia> @btime map!(xlogy, $out, $x, $y);
  780.819 μs (0 allocations: 0 bytes)

julia> @btime map!(logistic, $out, $x);
  717.576 μs (0 allocations: 0 bytes)

julia> @btime map!(logaddexp, $out, $x, $y);
  2.507 ms (0 allocations: 0 bytes)

julia> @btime sum(xlogx, $x);
  762.913 μs (0 allocations: 0 bytes)

[nalimilan]

Thanks! AFAIK with ? LLVM will decide whether to evaluate all branches or not, while with ifelse it will always evaluate. So if we trust LLVM ? is generally better.