Add windows known issue with cuda

llama-burn/README.md

@@ -130,3 +130,16 @@ instruction-tuned model based on the Llama2 architecture and tokenizer.
 Based on your hardware and the model selected, the `wgpu` backend might not be able to successfully
 run the model due to the current memory management strategy. With `cuda` selected, the precision is
 set to `f32` due to compilation errors with `f16`.
+
+### Windows
+
+The `cuda` backend is [unable to find nvrtc lib](https://github.com/coreylowman/cudarc/issues/246):
+
+```
+Unable to find nvrtc lib under the names ["nvrtc", "nvrtc64", "nvrtc64_12", "nvrtc64_123", "nvrtc64_123_0", "nvrtc64_120_3", "nvrtc64_10"]. Please open GitHub issue.
+```
+
+This has been fixed in the latest `cudarc` release (used by our `cuda-jit` backend), which is
+currently used [on main](https://github.com/tracel-ai/burn). To circumvent the issue, feel free to
+modify the code and use the latest Burn dependency in your project instead of `0.14.0`. This should
+also allow you to use `f16` precision (compilation errors have been fixed since).