1.18.8: @grad_from_chainrules fails when a rule argument is named func ("function argument name not unique"), breaks ImplicitAD #334

Version: ReverseDiff 1.18.8 (also master e1bed86), Julia 1.12.5, Linux x86_64. 1.18.7 works.

Since 1.18.8 (#328), @grad_from_chainrules generates methods whose own self-argument is named func:

(func::ReverseDiff._selftype(f))(args...; kwargs...) where {...} = ReverseDiff.track(func, ...)
function ReverseDiff.track(func::ReverseDiff._selftype(f), args...; kwargs...) where {...}

If the rule's signature has an argument that is itself named func, the generated signature has two arguments
called func and lowering fails:

using ReverseDiff, ChainRulesCore
k(func, x) = func(sum(x))
ReverseDiff.@grad_from_chainrules k(func, x::ReverseDiff.TrackedArray)
# ERROR: syntax: function argument name not unique: "func" around .../ReverseDiff/src/macros.jl:350

This breaks precompilation of ImplicitAD.jl, whose
src/external.jl:199 declares ReverseDiff.@grad_from_chainrules _provide_rule(func, x::ReverseDiff.TrackedArray, p, mode, jacobian, jvp, vjp),
and with it CCBlade.jl, FLOWVLM and FLOWUnsteady.

Possible fix: name the self-argument with a gensym, e.g. self = gensym(:func) and use $self in both
generated methods (lines 367 and 369 on master), so it cannot collide with a user-chosen name. The @grad
path (line 209) has the same pattern and probably the same issue for an argument named func; we have not
tested it.

We have not run ReverseDiff's test suite with a change; this is a report only.