make Gradient/Jacobian/Hessian-Config mutable

[KristofferC]

these often get quite large and the fact that they are non-mutable means Julia copies them which is quite expensive

Consider for example:

using ForwardDiff: ForwardDiff, HessianConfig, Chunk
using DiffResults

function test()
    x = rand(12)
    result = DiffResults.HessianResult(x)
    cfg = HessianConfig(sum, result, x, Chunk{12}())

    # Warmup
    ForwardDiff.hessian!(result, sum, x, cfg)

    @time ForwardDiff.hessian!(result, sum, x, cfg)

    @time for i in 1:100
        ForwardDiff.hessian!(result, sum, x, cfg)
    end
end

test()

Before this change it gives:

  0.000003 seconds (2 allocations: 16.000 KiB)
  0.000280 seconds (200 allocations: 1.562 MiB)

Now it instead gives

  0.000002 seconds (2 allocations: 160 bytes)
  0.000120 seconds (200 allocations: 15.625 KiB)

At some point, Julia might do this better.