feat: add whisper.cpp ROCm backend support for AMD GPU acceleration

- transcription.py: new _transcribe_remote_whispercpp() using /inference endpoint
- transcription.py: backend param routes to openai or whispercpp remote path
- config.py: whisper.backend default 'openai', alt 'whispercpp'
- pipeline.py: passes backend from config to transcribe_file
- settings: backend dropdown (OpenAI-compat / whisper.cpp)
- SETUP.md: whisper.cpp ROCm build and systemd setup instructions

whisper-cpp-server running on beastix :8080 (ROCm0, gfx1030, RX 6800 XT)

docs/SETUP.md

@@ -20,34 +20,41 @@ Einstellungsseite.
 
 ## Beastix (Server-Setup, einmalig)
 
-### 1. faster-whisper-server installieren
+### 1. whisper.cpp mit ROCm/GPU kompilieren
+
+Voraussetzung: ROCm installiert (Arch: `sudo pacman -S rocm-hip-sdk`).
 
 ```bash
-sudo pacman -S python-pipx   # Arch Linux
-pipx install faster-whisper-server
-pipx ensurepath
+mkdir -p ~/src && cd ~/src
+git clone https://github.com/ggml-org/whisper.cpp.git --depth=1
+cd whisper.cpp
+
+# Für AMD RX 6800 XT (gfx1030) — gfx-Target ggf. anpassen
+cmake -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release -DWHISPER_BUILD_SERVER=ON
+cmake --build build -j$(nproc)
+
+# Modell large-v3 herunterladen (~2.9 GB)
+bash models/download-ggml-model.sh large-v3
 ```
 
-**Bekannter Bug in Version 0.0.2** — fehlende `pyproject.toml` im pipx-venv:
-
-```bash
-cat > ~/.local/share/pipx/venvs/faster-whisper-server/lib/python*/site-packages/pyproject.toml << 'EOF'
-[project]
-name = "faster-whisper-server"
-version = "0.0.2"
-EOF
-```
+`gfx1030` = RX 6800 XT. Andere AMD GPUs: `rocminfo | grep gfx`
 
 ### 2. Als systemd-User-Service einrichten
 
 ```bash
-cat > ~/.config/systemd/user/faster-whisper-server.service << 'EOF'
+cat > ~/.config/systemd/user/whisper-cpp-server.service << 'EOF'
 [Unit]
-Description=faster-whisper-server (OpenAI-compatible Whisper API)
+Description=whisper.cpp Server (ROCm/GPU)
 After=network.target
 
 [Service]
-ExecStart=%h/.local/bin/faster-whisper-server --host 0.0.0.0 --port 8000 --model large-v3
+ExecStart=%h/src/whisper.cpp/build/bin/whisper-server \
+    --host 0.0.0.0 \
+    --port 8080 \
+    --model %h/src/whisper.cpp/models/ggml-large-v3.bin \
+    --language de \
+    --threads 4 \
+    --convert
 Restart=on-failure
 RestartSec=5
 
@@ -56,9 +63,12 @@ WantedBy=default.target
 EOF
 
 systemctl --user daemon-reload
-systemctl --user enable --now faster-whisper-server.service
+systemctl --user enable --now whisper-cpp-server.service
 ```
 
+Logs prüfen: `journalctl --user -u whisper-cpp-server -f`
+GPU-Nutzung bestätigt wenn in den Logs steht: `using ROCm0 backend`
+
 ### 3. Ollama installieren (falls noch nicht vorhanden)
 
 ```bash
@@ -105,7 +115,8 @@ Als Admin einloggen → Zahnrad-Icon im Header → Einstellungen:
 
 | Feld | Wert (Beispiel) |
 |------|-----------------|
-| Whisper Server URL | `http://beastix:8000` |
+| Whisper Backend | `whisper.cpp Server` |
+| Whisper Server URL | `http://beastix:8080` |
 | Whisper Modell | `large-v3` |
 | Ollama Server URL | `http://beastix:11434` |
 | Ollama Modell | `gemma3:12b` (aus Dropdown wählen) |