NaN parameters in dict_list change @onlyif behaviour

[lhupe]

By some spooky action at a distance, introducing a NaN-valued parameter into a dict_list changes the behaviour of @onlyif restrictions, even if they shouldn't depend on that parameter at all:

Minimal Working Example
Versions: DrWatson v2.9.1 and Julia v1.7.1

dict_list(Dict(
    :a => Inf,
    :b => [@onlyif(:c, 1), 2],
    :c => [true, false]
))

returns

3-element Vector{Dict{Symbol, Real}}:
 Dict(:a => Inf, :b => 1, :c => true)
 Dict(:a => Inf, :b => 2, :c => true)
 Dict(:a => Inf, :b => 2, :c => false)

while

dict_list(Dict(
    :a => NaN,
    :b => [@onlyif(:c, 1), 2],
    :c => [true, false]
))

returns

4-element Vector{Dict{Symbol, Real}}:
 Dict(:a => NaN, :b => 1, :c => true)
 Dict(:a => NaN, :b => 2, :c => true)
 Dict(:a => NaN, :c => false)
 Dict(:a => NaN, :b => 2, :c => false)

I've had a quick look at the dict_list code and spotted the problem: the last pass on the trial parameter combinations filters out all combinations that are subsets of other parameter combinations. Normally, this should remove the extra entry in the MWE since it only differs from the last entry by the missing dependent parameter, however, since NaN != NaN, Dict(:a => NaN, :c => false) is technically not a subset of Dict(:a => NaN, :b => 2, :c => false) and the comparison in dict_list.jl#92 fails.

My initial idea for a fix is to replace the trials[k] == _trials[k] with trials[k] == _trials[k] || trials[k] === _trials[k], since the === catches the NaNs, but the == is still necessary to catch mutable objects of equal value. Note that I have not thought this through at all, there may well be some edge case which breaks this.

[Datseris]

sounds good, maybe open a PR with the === check? Also cc @JonasIsensee maybe he knows more. actually, I don't remember anymore who contribute the @onlyif macro...