`(Mutable)DiffResult` allocates because it's a mutable struct

[gdalle]

That probably explains the one allocation we see even with optimal config on ForwardDiff and ReverseDiff

[gdalle]

The answer might be to include it in extras

[gdalle]

JuliaDiff/DiffResults.jl#26

[gdalle]

No improvement possible because we cannot put it in extras. Otherwise future gradient calls might modify something we have already returned