Subtype: Avoid split y::Union when comparing with a forall var.

[N5N3]

Should bring the load time of CUDA.jl back to a similar level before the regression.
MWE:

julia> @time !<:(Tuple{Pair{<:T,<:T}, Val{S} where {S}} where {T<:Base.BitInteger},
                 Tuple{Pair{<:T,<:T}, Val{Int}} where {T<:Base.BitInteger})
  0.000016 seconds (33 allocations: 2.188 KiB) # master 86.345139 seconds (50.49 M allocations: 7.023 GiB, 0.53% gc time)

close #48582.

[staticfloat]

Thanks, @N5N3!