optimize permute! and invpermute!

[piever]

After JuliaLang/julia#44941 permute! and invpermute! no longer call the Base.permute!! and Base.invpermute!! methods. This ensures that permute! and invpermute! stay optimized for PooledArrays, otherwise they get a suboptimal fallback (see benchmark below).

julia> using BenchmarkTools, PooledArrays, Random

julia> perm = randperm(10_000);

julia> @btime permute!($v, $perm) setup=(v=PooledArray(rand(1:1000, 10_000)));
  9.800 μs (2 allocations: 39.11 KiB)

julia> @btime ($v .= $v[$perm]) setup=(v=PooledArray(rand(1:1000, 10_000))); # current permute! fallback
  10.800 μs (4 allocations: 39.17 KiB)

julia> @btime invpermute!($v, $perm) setup=(v=PooledArray(rand(1:1000, 10_000)));
  11.900 μs (2 allocations: 39.11 KiB)

julia> @btime ($v[$perm] = $v) setup=(v=PooledArray(rand(1:1000, 10_000))); # current invpermute! fallback
  103.400 μs (4 allocations: 39.17 KiB)

The improvement is much more substantial for invpermute!, but I imagine it makes sense to have both methods fully optimized. I was actually surprised that the permute! fallback was already almost optimal. EDIT: that was just due to it calling the optimized getindex(v::PooledArray, i::AbstractVector) and copyto!(dst::PooledArray, src::PooledArray) methods, whereas invpermute! has to call setindex!, which is more costly.