test multiple evaluators: `findlast` over the evaluation-data `Dict` can return the wrong evaluator

power_flow_evaluations only ever registers one evaluator (src/core/interfaces.jl:187-191), and no test registers more than one. The multi-evaluator path is untested, and it doesn't look correct.

latest_solved_power_flow_evaluation_data (ext/PowerFlowsExt/pf_solve_and_aux.jl:7) does findlast(x -> x.is_solved, collect(values(...))) over a Dict{DataType, AbstractEvaluationData} (IOM src/core/external_evaluation.jl:35-38). A Dict has no "latest" — values iterates in hash-slot order over objectid-hashed DataType keys, stable within a session but not across sessions or Julia versions. With more than one entry, which evaluator comes back is arbitrary.

That matters because IOM.calculate_aux_variables! (IOM src/core/optimization_container.jl:1328-1339) resets all evaluators, then for each one calls evaluate! and re-runs every evaluator-bound aux var key through the 3-arg calculate_aux_variable_value!, which throws away the data the loop is holding and re-derives it via findlast, then filters on ownership (pf_solve_and_aux.jl:158-162):

• After the first evaluate!, only that evaluator is solved, so findlast returns it by luck.
• After the second, both are solved and findlast may return the first. The second evaluator's keys then hit the || return and are silently never written — or, if both expose the same aux var type, are written from the wrong power flow.

No error, no warning: same silent-failure shape as #305.

Direction: have the IOM loop pass data to the existing 4-arg calculate_aux_variable_value! (pf_solve_and_aux.jl:95, :114) with the ownership filter moved onto a 4-arg POM method, dropping findlast from the path. Touches IOM too, so separate from #305.

Tests to add: two evaluators with disjoint aux var sets (all populated); two exposing the same aux var type (value comes from the intended one). Failure depends on Dict ordering, so a single green run proves little.

Related: #305. Also, my comment on #273 calls this findlast "the most recent simulation step" — wrong: evaluation_data is keyed by evaluator type, and reset_evaluations! clears the flags at the start of every solve, so nothing survives across steps.