fix unsoundness in fieldtype of Tuple with non-Type params

[JeffBezanson]

This is one way to fix #39988. The alternative is to make fieldtype(Tuple{:x}, 1) return Union{}, which would be more consistent with other types. But, I'm assuming the default right thing is to make inference correct rather than change the value we return, so I'm trying this first. This version would be backportable for instance. But, if it causes problems (e.g. maybe too many new invalidations) we could just make the change to fieldtype instead and give up on fixing it for <1.7.

[vtjnash]

Regardless of what we decide here for now, I believe fieldtype and first-parameter may eventually need to diverge somewhat to fix this mistake with the diagonal rule / covariance:

julia> Tuple{Tuple{T, T}, Val{T}} where T
Tuple{Tuple{T, T}, Val{T}} where T

julia> fieldtype(ans, 1)
Tuple{T, T} where T # wrong: should be `Tuple{Any, Any}`

julia> ((1, 2.), Val{Any}()) isa Tuple{Tuple{T, T}, Val{T}} where T
true

julia> getfield(((1, 2.), Val{Any}()), 1) isa fieldtype(Tuple{Tuple{T, T}, Val{T}} where T, 1)
false

[maleadt]

The server has been rebooted, so let's try once more:

@nanosoldier runtests(ALL, vs = ":master")

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected. A full report can be found here. cc @maleadt

[JeffBezanson]

Looks like all the failures are numeric issues, benchmark noise, etc.

[KristofferC]

Could this be rebased on #40702? I am not sure what the replacements for jl_is_vararg and jl_vararg_t are.