Consider making multiply more explicit WRT overflow

Currently `a:uN * b:uN` is converted to IR `uN = umul[a, b]`. This means that for some values of `a` and `b` overflow is possible.

We might want to consider desuggaring `a*b` to a umul of 2N width and then truncating to N bits. This is more explicit about whats actually going on.

We will need to consider the benefits and costs associated with this.