affine_transform + broadcasting = segfault

Consider the following code:

```
import dynet as dy

W = dy.zeros((32, 32))
x = dy.zeros((32,))
b = dy.zeros((32,64))

y = W * x + b
print(y.dim())

z = dy.affine_transform([b, W, x])
print(z.dim())

z.forward()
z.backward()
```

Here the dimensionality of `y` is correct (`((32, 64), 1)`), but the dimensionality of `z` is incorrect (dynet thinks it's `((32,), 1)` when it should be the same as `y`).

Furthermore, calling `backward()` on `z` causes a segfault.

Start by running the provided C++-library reproducer and inspect the affine_transform entry point and the broadcasting behavior used by W * x + b. Confirm that affine_transform reports the broadcasted dimensions, then run forward and backward to verify that the corrected result no longer segfaults.

Written by the indexing model from the issue text.