Optimize conversions duplicated across phi nodes

In the case that the IR contains patterns such as "Phi(ConvX2Y(A), ConvX2Y(B))", the convert that is duplicated can be pulled after the phi. This would lead to the resultant IR "ConvX2Y(Phi(A, B))" which is more compact.

This pattern is fairly widespread as it is fairly easy to generate in user code. It's also generated when a new array with a non-constant size is created, so this transformation would be pretty widely applicable.