arraytable significantly slows down when passed table with heterogenous columns

Here is the test code:

using JSONTables, DataFrames

results = DataFrame(cols=Int[], rows=Int[], arraytable=Float64[], objecttable=Float64[])
for cols in (100, 200, 300, 400, 500), rows in (10^3, 10^4, 10^5)
    @show (cols, rows)
    df = DataFrame(ones(rows, cols))
    df[!, 1] .= "a"
    df[!, 2] .= 1
    df[!, 3] .= true
    open(io -> objecttable(io, df), "test.json", "w")
    x1 = @elapsed open(io -> arraytable(io, df), "test.json", "w")
    open(io -> objecttable(io, df), "test.json", "w")
    x2 = @elapsed open(io -> objecttable(io, df), "test.json", "w")
    push!(results, [cols, rows, x1, x2])
end

and here is the benchmark result:

julia> results
15×4 DataFrame
│ Row │ cols  │ rows   │ arraytable │ objecttable │
│     │ Int64 │ Int64  │ Float64    │ Float64     │
├─────┼───────┼────────┼────────────┼─────────────┤
│ 1   │ 100   │ 1000   │ 0.178669   │ 0.0328257   │
│ 2   │ 100   │ 10000  │ 1.65927    │ 0.26272     │
│ 3   │ 100   │ 100000 │ 16.2332    │ 2.36529     │
│ 4   │ 200   │ 1000   │ 0.349468   │ 0.0498271   │
│ 5   │ 200   │ 10000  │ 3.49195    │ 0.595482    │
│ 6   │ 200   │ 100000 │ 34.8853    │ 4.9485      │
│ 7   │ 300   │ 1000   │ 0.547324   │ 0.0803132   │
│ 8   │ 300   │ 10000  │ 5.16746    │ 0.759614    │
│ 9   │ 300   │ 100000 │ 52.3498    │ 7.48296     │
│ 10  │ 400   │ 1000   │ 0.714898   │ 0.104794    │
│ 11  │ 400   │ 10000  │ 6.91257    │ 1.00389     │
│ 12  │ 400   │ 100000 │ 73.8235    │ 11.4878     │
│ 13  │ 500   │ 1000   │ 0.947894   │ 0.146453    │
│ 14  │ 500   │ 10000  │ 10.5129    │ 1.42812     │
│ 15  │ 500   │ 100000 │ 94.0811    │ 13.1131     │

Start by running the provided Julia benchmark and compare the arraytable and objecttable entry points on heterogeneous DataFrames. Trace the arraytable path to identify why its runtime grows so much faster, then verify the change against the same combinations of columns and rows. Done means arraytable handles heterogeneous columns without the reported performance gap.