build(cuda): ship a CUDA installer variant with bundled runtime DLLs

The default build (default = candle-cuda) load-time-imports the CUDA runtime,
so it crashes on any machine without the toolkit. Add a CUDA release variant
that bundles the redistributable runtime DLLs (cudart/cublas/cublasLt/curand,
verified via dumpbin) so it runs on any NVIDIA machine with just a driver.

- tauri.cuda.conf.json: a 'tauri build --config' overlay (CPU release
  untouched) that points the updater at latest-cuda.json so a CUDA install
  never self-updates into the CPU build
- windows/cuda-hooks.nsh: NSIS hook embeds the DLLs via File and extracts them
  next to phokus.exe, where the Windows loader searches
- build:app:cuda script; cuda-redist/ gitignored (DLLs copied from the toolkit
  locally), with a negation so the overlay stays tracked

.gitignore

@@ -35,3 +35,10 @@ dist-ssr
 *.json
 
 *.pyc
+
+# Bundled CUDA runtime DLLs for the CUDA build (copied from the toolkit
+# locally; ~600 MB, never committed). See RELEASE_PLAN.md "CUDA release variant".
+src-tauri/cuda-redist/
+
+# Keep the CUDA build config overlay tracked despite the *.json rule above.
+!src-tauri/tauri.cuda.conf.json