Support for Nvidia unified memory?

All,

I saw this morning that Tim Dettmers bitsandbytes python lib uses Nvidia's [Unified Memory](https://developer.nvidia.com/blog/unified-memory-cuda-beginners/) by [default](https://x.com/stasbekman/status/1749968490155696612), see (`csrc/pythonInterface.c:377`).

It doesn't look like candle, via cudarc supports this.

I'm interested in supporting this feature in Rust because my current GPU has a decent compute score, but only 4GB of VRAM (T1000).

Is there any plan to support this in candle?

Start by reading the bitsandbytes reference at csrc/pythonInterface.c:377 and tracing how candle currently reaches cudarc. Determine whether cudarc can expose Nvidia Unified Memory and what candle integration would require; done means a decided implementation path and verified support for the reported low-VRAM use case.

Written by the indexing model from the issue text.