`broadcast_as` error when processing multiple tokens at once in quantized example

Hello all,

Thanks for your great work here. We are implementing speculative decoding at mistral.rs, and were in the final stages of testing when we discovered some incredibly strange behavior. Specifically, the following error results when sending multiple tokens at once during the completions steps:

`Error: cannot broadcast [3, 3] to [1, 32, 3, 5]`

Reproducing this error is simple:

In the [`quantized/main.rs:578`](https://github.com/huggingface/candle/blob/a09d451d11a91ea7a7feaa40460abb282581a0f1/candle-examples/examples/quantized/main.rs#L578):

```diff
- let input = Tensor::new(&[next_token], &device)?.unsqueeze(0)?;
+ let input = Tensor::new(&[next_token, next_token, next_token], &device)?.unsqueeze(0)?;
```

Is this a bug?