Switch to ROCm 6.1 + torch 2.5.1 to fix MIOpen workspace=0 slowness

ROCm 7.2 + PyTorch 2.11.0 has a bug where PyTorch passes workspace=0 to
MIOpen convolutions, forcing fallback to the slow GemmFwdRest solver.
This caused s3gen.inference to take 15-22s instead of <5s, making
synthesis 3-4x slower than real-time audio playback.

ROCm 6.1 allocates workspace correctly so MIOpen picks fast GEMM solvers
without needing torch.compile workarounds.

Changes:
- Base image: rocm/dev-ubuntu-22.04:7.2 → 6.1
- torch 2.11.0 → 2.5.1 (rocm6.1 wheel index)
- Add pytorch_triton_rocm==3.1.0
- transformers 5.2.0 → 4.46.3, safetensors 0.5.3 → 0.4.0
- s3tokenizer unpinned → 0.3.0
- resemble-perth==1.0.1 directly (v1.0.1 is pip-installable; drop stub)
- Drop Dockerfile perth_stub steps
- Drop torch.compile and timing patches from engine.py (not needed)
- Drop multi-pass warmup from main.py (torch JIT warmup not needed)
- Drop ROCm 7.2-specific env vars from docker-compose.yml

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

main.py

@@ -18,33 +18,16 @@ logging.basicConfig(
 logger = logging.getLogger(__name__)
 
 
-_WARMUP_TEXTS = [
-    # Short: covers brief HA notifications (lights on/off, etc.)
-    "Okay.",
-    # Medium: covers typical HA announcements
-    "The front door is open. Please close it.",
-    # Long: covers longer TTS requests and pre-compiles dynamic shape graph
-    (
-        "This is a warmup synthesis to pre-compile neural network kernels "
-        "for longer text lengths used in Home Assistant announcements and notifications."
-    ),
-]
-
-
 def _warmup(voices: dict) -> None:
+    """Run one synthesis to populate MIOpen's in-memory kernel cache."""
     from wyoming_voices import resolve_voice
     audio_prompt = resolve_voice(None, voices) if voices else None
-    logger.info(
-        f"Running {len(_WARMUP_TEXTS)}-pass warmup to pre-compile torch kernels "
-        "for short, medium, and long text lengths..."
-    )
-    for i, text in enumerate(_WARMUP_TEXTS, 1):
-        try:
-            engine.synthesize(text=text, audio_prompt_path=audio_prompt)
-            logger.info(f"Warmup pass {i}/{len(_WARMUP_TEXTS)} complete")
-        except Exception:
-            logger.warning(f"Warmup pass {i} failed (non-fatal)", exc_info=True)
-    logger.info("Warmup complete")
+    logger.info("Running warmup synthesis...")
+    try:
+        engine.synthesize(text="Warmup.", audio_prompt_path=audio_prompt)
+        logger.info("Warmup complete")
+    except Exception:
+        logger.warning("Warmup synthesis failed (non-fatal)", exc_info=True)
 
 
 async def main() -> None:
@@ -58,10 +41,6 @@ async def main() -> None:
     voices = load_voices()
     wyoming_info = create_wyoming_info(engine.get_sample_rate(), voices)
 
-    # Run a warmup synthesis before accepting connections so MIOpen benchmarks
-    # and caches the best convolution algorithms for all layer shapes. Without
-    # this, the first real HA request triggers benchmarking (hundreds of runs)
-    # and times out before any audio is returned.
     _warmup(voices)
 
     host = get_wyoming_host()