[vm] Handle late fields in type propagation

[liamappelbe]

Currently we're just setting late fields to nullable in the type propagator, because otherwise all the sentinel comparison logic gets optimised out. This works, but is not optimal.

We'll need to extend CompileType to be able to handle lateness.

[mraleph]

I don't think you should extend CompileType for this purpose. I suggest instead look at unifying late field lowering and InitStaticField and then either write a specialized data flow pass which removes redundant Init* instructions (if we want to handle complicated cases) or just permit CSE to remove Init* by dominance.