Consider changing `BroadcastTo` to allow its shape to be derived from another symbol

Say you want to do
```python
x = model.set(5).reshape(-1, 1)
y = model.constant(np.arange(3).reshape(1, 3))
z = x + y
```
this should be valid because the resulting `z` should have shape `(-1, 3)`. But unfortunately our current `BinaryOp` implementation does not support non-scalar broadcasting and our `BroadcastTo` symbol does not support getting the desired shape from another symbol.

We should change the latter. Something like
```python
xb = broadcast_to(x, shape=(-1, 3)) # we can do this now
yb = broadcast_to(y, shape=x) # this saying `yb` should have the shape that would result from broadcasting `y` and `x`

```