Large number of allocations in 0.6.8

The eager automatic conversion introduced in #227 (cc @timholy @piever) leads to many allocations when this conversion isn't actually needed.
For example:
StructArrays@0.6.7 - 25 allocations:

julia> using StructArrays, ArraysOfArrays

julia> A = StructArray(x=VectorOfVectors{Int}())

julia> @time for _ in 1:10^6 push!(A, (x=1:10,)) end
  0.307580 seconds (25 allocations: 11.269 MiB, 7.45% gc time)

StructArrays@0.6.8 - 1 million allocations:

julia> using StructArrays, ArraysOfArrays

julia> A = StructArray(x=VectorOfVectors{Int}())

julia> @time for _ in 1:10^6 push!(A, (x=1:10,)) end
  0.809896 seconds (1.00 M allocations: 293.601 MiB, 55.02% gc time)

Of course, these allocations go away if automatic conversion (https://github.com/JuliaArrays/StructArrays.jl/blob/1581d70090de650646277e4eee4e33794e55b9ac/src/utils.jl#L197-L199) is disabled by StructArrays.maybe_convert_elt(::Type{T}, vals::Tuple) where {T} = vals.

This change in 0.6.8 is highly breaking performance-wise. Can the new automatic conversion approach be revised and made less eager somehow?

Start with the allocation benchmark in the issue, comparing StructArrays 0.6.7 and 0.6.8. Inspect the automatic-conversion logic in src/utils.jl at the referenced lines and the behavior of maybe_convert_elt(::Type{T}, vals::Tuple). Done means revising the conversion behavior so the reported push! workload no longer incurs one million allocations while required conversion still works.