@grad fails for qualified function names (gensym(::Expr)) #318

@grad fails at macro expansion time if the function name is qualified, e.g. M.f or Base.:*:

using ReverseDiff

module M
    f(x) = sum(abs2, x)
end

M.f(x::ReverseDiff.TrackedArray) = ReverseDiff.track(M.f, x)
ReverseDiff.@grad function M.f(x)
    xv = ReverseDiff.value(x)
    return M.f(xv), Δ -> (2 .* Δ .* xv,)
end

ERROR: LoadError: MethodError: no method matching gensym(::Expr)
Stacktrace:
 [1] var"@grad"(__source__::LineNumberNode, __module__::Module, expr::Any)

Expected: this defines the rule, like it does for an unqualified name. Qualified names are the normal way to add a rule for a function you don't own, e.g. a Base or LinearAlgebra function for your own array type. @grad_from_chainrules LinearAlgebra.dot(...) already handles them.

Cause: closure = gensym(f) assumes f is a Symbol. A plain gensym(:grad) (or gensym()) would be enough. This is the same line as #135 (callable structs), but a fix for that issue alone won't necessarily cover this case.

ReverseDiff master (b796032, v1.18.4), Julia 1.13.1.