Zero device memory allocation tensor ops support

Is there any way to do tensor ops(add/div/matmul...) without memory allocation?
eg:
```
let dst = Tensor::zeros((2, 2), DType::F32, &cuda_dev)?;
let lhs= Tensor::zeros((2, 2), DType::F32, &cuda_dev)?;
let rhs= Tensor::zeros((2, 2), DType::F32, &cuda_dev)?;
candle::add(&lhs,&rhs, &mut dst)?; //zero device memory allocation here
```

No source files or tests are identified. Start by tracing CUDA tensor creation and the candle::add operation, then determine how add, div, and matmul could write into a provided destination without device allocation; done means these operations demonstrate zero additional device-memory allocation.

Written by the indexing model from the issue text.