Write CUDA kernels in Rust

Hi there, I wanna bring this up that https://github.com/NVlabs/cuda-oxide has been actively developed, and the way of writing kernels in `.cu` can be ported to Rust.

Start by reviewing the linked cuda-oxide project and Candle's existing kernel approach, comparing the current .cu workflow with the proposed Rust workflow. Before coding, define which kernels and integration points are in scope; done should include a working Rust kernel path in Candle and evidence that the relevant existing behavior is preserved.

Written by the indexing model from the issue text.