Fix torch.compile crash on torch.device ConstantVariable bug

The installed PyTorch can't treat torch.device as a ConstantVariable
during tracing (hit in diffusers' get_timestep_embedding). Setting
suppress_errors=True converts the hard crash to a graph break so dynamo
falls back to eager for that subgraph while still compiling the 60
transformer blocks.

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

app.py

@@ -95,6 +95,7 @@ try:
 except Exception as _fa3_err:
     print(f"[startup] FA3 not available ({_fa3_err}), using default attention processor.", flush=True)
 
+torch._dynamo.config.suppress_errors = True
 print("[startup] applying torch.compile to transformer (mode=reduce-overhead)...", flush=True)
 pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead")
 print("[startup] torch.compile done.", flush=True)