Handle PhiNode with edge==0

[martinholters]

I'm not entirely sure this makes sense. Maybe we shouldn't produce PhiNodes with an edge of 0 in the first place (inserting a dummy instruction as first BB before it instead or something). But I do get the same @code_llvm for f() = while true x = try finally end end and g() = (nothing; while true x = try finally end end), where before, inference of f() would give an internal error. Also, I note that once upon a time, we used to do something similar:

julia/base/compiler/ssair/legacy.jl

Lines 197 to 201 in 61b75ef

	if isa(val, PhiNode)
	# Translate from BB edges to statement edges
	edges = Any[edge == 0 ? 0 : terminator_mapping[edge] for edge in val.edges]
	@assert all(x->x >= 0, edges)
	val = PhiNode(convert(Vector{Any}, edges), copy(val.values))

I'm not sure a 0-edge might cause trouble elsewhere, though. But then again, it would not have worked at all before and did not seem to cause too much trouble, so those PhiNodes seem to be quite rare, and at least the case at hand is fixed by this.

Fixes #45600.