bug: index_select() does not work on CUDA if index array has zero elements

Here's a minimal reproduction:

```rust
use candle_core::{Device, Tensor};

pub fn main() -> candle_core::Result<()> {
let device = Device::Cpu;
let indices = Tensor::from_vec::<_, u32>(vec![], 0, &device)?;
let sample = Tensor::randn(0., 1., (3, 3, 3), &device)?;

// works fine, produces a result with shape (0, 3, 3)
let result = sample.index_select(&indices, 0)?;
dbg!(result);

let device = Device::new_cuda(0)?;
let indices = Tensor::from_vec::<_, u32>(vec![], 0, &device)?;
let sample = Tensor::randn(0., 1., (3, 3, 3), &device)?;

// crashes with Cuda(Cuda(DriverError(CUDA_ERROR_INVALID_VALUE, "invalid argument")))
let result = sample.index_select(&indices, 0)?;
dbg!(result);

Ok(())
}
```

`candle-core` version is 0.8.1

Start by reproducing the issue with the Rust example in the report, then trace the CUDA implementation of Tensor::index_select and its handling of zero-length index tensors. Compare it with the CPU behavior and run the relevant CUDA tests or add coverage for an empty index array; done means the operation returns a tensor shaped (0, 3, 3) without a CUDA error.

Written by the indexing model from the issue text.