Codegen nondeterminism based on inference cache (emit_invoke)

[maleadt]

In a similar vein to #35800, I ran into emit_invoke codegen depending on the state of the inference cache:

; julia> @noinline child(@nospecialize(i)) = i

; julia> kernel(a) = (child(a); return)

; julia> code_llvm(kernel, Tuple{Int}; optimize=false)

;  @ REPL[2]:1 within `kernel'
define void @julia_kernel_94(i64) {
top:
  %1 = call %jl_value_t*** @julia.ptls_states()
  %2 = bitcast %jl_value_t*** %1 to %jl_value_t**
  %3 = getelementptr inbounds %jl_value_t*, %jl_value_t** %2, i64 4
  %4 = bitcast %jl_value_t** %3 to i64**
  %5 = load i64*, i64** %4
  %6 = call %jl_value_t* @jl_box_int64(i64 signext %0)
  %7 = call cc38 nonnull %jl_value_t* bitcast (%jl_value_t* (%jl_value_t*, %jl_value_t**, i32, %jl_value_t*)* @jl_invoke to %jl_value_t* (%jl_value_t*, %jl_value_t*, %jl_value_t*)*)(%jl_value_t* inttoptr (i64 139913193714448 to %jl_value_t*), %jl_value_t* inttoptr (i64 139913178153096 to %jl_value_t*), %jl_value_t* %6)
  ret void
}

; julia> code_llvm(child, Tuple{Any}; optimize=false); # populate the cache

; julia> code_llvm(kernel, Tuple{Int}; optimize=false)

;  @ REPL[2]:1 within `kernel'
define void @julia_kernel_159(i64) {
top:
  %1 = call %jl_value_t*** @julia.ptls_states()
  %2 = bitcast %jl_value_t*** %1 to %jl_value_t**
  %3 = getelementptr inbounds %jl_value_t*, %jl_value_t** %2, i64 4
  %4 = bitcast %jl_value_t** %3 to i64**
  %5 = load i64*, i64** %4
  %6 = call %jl_value_t* @jl_box_int64(i64 signext %0)
  %7 = call cc37 nonnull %jl_value_t* bitcast (%jl_value_t* (%jl_value_t*, %jl_value_t**, i32)* @j1_child_160 to %jl_value_t* (%jl_value_t*, %jl_value_t*)*)(%jl_value_t* inttoptr (i64 139913178153096 to %jl_value_t*), %jl_value_t* %6)
  ret void
}

Looks like this was added in #25984. I'm not sure it's wanted / worth it?