Ensure we propagate the return index when a return block is folded into an existing block

[amfaber]

Ensure we propagate the return index when a return block is folded into an existing block

Without this change, my trivial program that I scaled down to in order to debug something failed to compile.
The trivial program was something like this

#[cube(launch_unchecked)]
pub fn minimal_reproduction(
    mut output: Tensor<Line<f32>>,
){
    let row = ABSOLUTE_POS;
    let rows = output.shape(0);
    let cols = output.shape(1);
    if row >= rows {
        terminate!()
    }
}

During cubecl's optimization passes, this would eventually look for post dominators during GVN

impl Analysis for PostDominators {
    fn init(opt: &mut crate::Optimizer) -> Self {
        let mut reversed = opt.program.graph.clone();
        reversed.reverse();
        PostDominators(dominators::simple_fast(&reversed, opt.ret))
    }
}

Since the whole program had been folded into a single basic block, but the opt.ret index had not been updated, this would attempt to traverse the graph starting from a non-existent index, leading to a panic.

This change ensures that the opt.ret index is updated if we remove the block it was previously pointing to, at least in this pass.

[wingertge]

Should put this in update_references so it applies to all merge passes. I did indeed forget about the that value 😅