Fix spurious allocations in ret_allocs test helper on Julia 1.10/1.11

[ChrisRackauckas-Claude]

Summary

The "Non-allocating cache construction" testset (test/finitedifftests.jl) currently fails on Julia 1.10 and 1.11 with 12 Evaluated: 48 == 0 failures at :371/:379. It passes on Julia 1.12. This is a test-helper artifact, not a library regression — the finite_difference_gradient! path is genuinely allocation-free.

Root cause

function ret_allocs(res, _f, x, cache)
    @allocated FiniteDiff.finite_difference_gradient!(res, _f, x, cache)
end

Julia's default heuristic does not specialize a method on a Function-typed argument that is only forwarded to an inner call. So _f is boxed inside ret_allocs, the inner finite_difference_gradient! becomes a dynamic dispatch, and the wrapper itself allocates ~48 bytes on 1.10/1.11 (0 on 1.12+, where the heuristic improved). The library work is 0 bytes.

Independently verified the library allocates nothing:

• @allocated inside a function around the call = 0
• 1000 direct calls measured via Base.gc_bytes() = 0 total
• Profile.Allocs on a normally-compiled call = 0 allocations

Fix

Force specialization on the function argument:

function ret_allocs(res, _f::F, x, cache) where {F}
    @allocated FiniteDiff.finite_difference_gradient!(res, _f, x, cache)
end

Verification

Ran the full Core suite locally on Julia 1.11: FiniteDiff Standard Tests | 219 pass, 0 fail, 1 broken → Testing FiniteDiff tests passed (previously 207 pass / 12 fail / 1 broken).

Please ignore until reviewed by @ChrisRackauckas.

🤖 Generated with Claude Code