`@batch` uses only the first 64 threads

I am currently working with a 128 cores (64 cores x 2 sockets) CPU, and I noticed that Polyester.jl will only use the first 64 threads available :

```julia
julia> using Polyester

julia> Threads.nthreads()
128

julia> tids = zeros(Int, Threads.nthreads());

julia> @batch for _ in 1:Threads.nthreads()
tids[Threads.threadid()] += 1
end

julia> all(tids .== 1)
false

julia> [tids[1:64]' ; tids[65:end]']
2x64 Matrix{Int64}
2 2 2 2 2 2 2 2 ... 2 2 2 2 2 2 2 2
0 0 0 0 0 0 0 0 ... 0 0 0 0 0 0 0 0

julia> tids .= 0;

julia> Threads.@threads for _ in Threads.nthreads()
tids[Threads.threadid()] += 1
end

julia> all(tids .== 1)
true

julia> [tids[1:64]' ; tids[65:end]']
2x64 Matrix{Int64}
1 1 1 1 1 1 1 1 ... 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 ... 1 1 1 1 1 1 1 1
```

Notice how the threads 65 to 128 were ignored by `@batch` but not by `Threads.@threads`.
I get similar results with hyperthreading with 256 threads.

I am quite sure that this also affects LoopVectorization.jl since I am getting the same time for some very simple benchmarks :
```julia
julia> a = rand(Float64, 10000); b = rand(Float64, 10000); c = rand(Float64, 10000);

julia> @btime @tturbo for i in 1:10000
c[i] = a[i] * b[i]
end
```

Gives ~3.0 µs for 64 or 128 threads.

I suppose that it is related to the behaviors mentioned in #22. In any case I would be happy to help resolving this issue.