Tensor::arange allocates first on the CPU even when the target device is a GPU

`Tensor::arange` calls `Tensor::arange_step` which works by pushing values to a `Vec` stored on the CPU. I think this could be more efficient (firstly by precomputing the length and using `Vec::with_capacity`), and second to make the tensor on the device directly if it is supposed to end up on the GPU.

A parallel algorithm could be to use the thread index of the CUDA thread to compute the value `tensor[idx] = start + step * idx`. You could theoretically also do this on the CPU with something like rayon, but probably only for large tensors.

Start with Tensor::arange and its arange_step implementation, then trace how values are collected into the CPU Vec and transferred to the target device. Compare the proposed preallocated collection and CUDA thread-index approach, and consider the CPU path before verifying that arange produces the expected values directly on the requested device.

Written by the indexing model from the issue text.