nvidia ux & bookworm bpo cleanup

- Removed redundant confirmation dialogs from `_install_nvidia_stack`. Users can now access driver management (including removal) without first confirming installation intent; initial plan display is followed by pause for review instead of blocking prompt.
- Automated CUDA repository enabling on Debian 13/Bookworm paths (590/595). Manual confirmation prompts eliminated; script auto-enables repo via `_enable_cuda_repo` with error handling while maintaining safety checks.
- Split CUDA driver installation into two distinct APT transactions in `nvidia.sh`. Step 1 installs mandatory version pinning (`nvidia-driver-pinning-${ver}`); failure aborts process. Step 2 installs `nvidia-open`, relying on transitive dependencies for firmware rather than explicit package listing.
- Eliminated all Bookworm backports support due to EOL (2026-08-09). Removed `_install_nvidia_bookworm_bpo()`, dead code `install_nvidia_driver()`, and dispatcher logic for Debian 12. Turing now installs standard packages only; Trixie backports remain intact.
- Updated documentation (`docs/gpu.md`) to reflect actual installation flow and removed references to deprecated functions.

modules/utils.sh

@@ -492,6 +492,9 @@ get_intel_generation() {
 
 # ----------------------------------------------------------------------
 # Check if backports repository is enabled (active line without #)
+# Nota: el flujo NVIDIA Debian 12 (Bookworm) ya NO usa esta función
+# (backports EOL 2026-08-09). Sigue activa para kernels, desktop/greetd,
+# gaming, audio, comunicación y Mesa (AMD/Intel).
 # ----------------------------------------------------------------------
 is_backports_enabled() {
     local codename="${DEBIAN_CODENAME:-}"