Inference: propagate struct initialization info on setfield!

[serenity4]

When a variable has a field set with setfield!(var, field, value), inference now assumes that this specific field is defined and may for example constant-propagate isdefined(var, field) as true. PartialStruct, the lattice element used to encode this information, still has a few limitations in terms of what it may represent (it cannot represent mutable structs with non-contiguously defined fields yet), further work on extending it would increase the impact of this change.

Consider the following function:

julia> function f()
           a = A(1)
           setfield!(a, :y, 2)
           invokelatest(identity, a)
           isdefined(a, :y) && return 1.0
           a
       end
f (generic function with 1 method)

Here is before on master:

julia> @code_typed f()
CodeInfo(
1 ─ %1 = %new(Main.A, 1)::A
│          builtin Main.setfield!(%1, :y, 2)::Int64
│        dynamic builtin (Core._call_latest)(identity, %1)::Any
│   %4 =   builtin Main.isdefined(%1, :y)::Bool
└──      goto #3 if not %4
2 ─      return 1.0
3 ─      return %1
) => Union{Float64, A}

And after this PR:

julia> @code_typed f()
CodeInfo(
1 ─ %1 = %new(Main.A, 1)::A
│          builtin Main.setfield!(%1, :y, 2)::Int64
│        dynamic builtin (Core._call_latest)(identity, %1)::Any
└──      return 1.0
) => Float64

[serenity4]

The mechanism used to propagate PartialStruct refinement is frame-local however, and will currently not propagate if using setproperty! instead of setfield! in the example above (because setfield! is then inferred in the context of setproperty! and the refined type information is not saved once done with setproperty!).

[serenity4]

This should be ready (if no new issues are detected in tests) @aviatesk

[aviatesk]

LGTM.

[aviatesk]

Restarted test revise. Once confirming it passes successfully, we can merge this.