Hi,

I am trying to do in-place random number generation using `Distributions.jl`

. However, for some reason, only the random number generation from the Categorical distribution takes extra computation time due to the extra allocations. Here is the code.

```
function loop()
w = rand(10^5)
normalize!(w,1)
tmpI = zeros(Int,10^5)
tmpF = zeros(Float64,10^5)
@time for i in 1:10^3
rand!(Categorical(w), tmpI)
end
@time for i in 1:10^3
rand!(Normal(0,1), tmpF)
end
@time for i in 1:10^3
rand!(Gamma(1,1), tmpF)
end
@time for i in 1:10^3
rand!(InverseGamma(1,1), tmpF)
end
@time for i in 1:10^3
rand!(Uniform(0,1), tmpF)
end
end
loop()
```

And here is the result.

```
4.153208 seconds (10.00 k allocations: 3.726 GiB, 1.42% gc time)
0.269260 seconds
0.444689 seconds
0.883855 seconds
0.063641 seconds
```

Is this inevitable?