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>

requirements-rocm.txt

@@ -5,18 +5,18 @@ librosa==0.11.0
 pyloudnorm
 
 # ML dependencies (pinned to match chatterbox without overwriting ROCm torch)
-transformers==5.2.0
+transformers==4.46.3
 diffusers==0.29.0
-safetensors==0.5.3
+safetensors==0.4.0
 huggingface-hub
 omegaconf
 
 # Chatterbox dependencies (installed separately since chatterbox uses --no-deps)
-# Note: resemble-perth is stubbed out in perth_stub.py (watermarking unneeded for self-hosted use)
 conformer==0.3.2
-s3tokenizer
+s3tokenizer==0.3.0
 spacy-pkuseg
 pykakasi==2.3.0
+resemble-perth==1.0.1
 
 # Wyoming protocol
 wyoming>=1.5.4