demo.py: cap diffuse n_steps_list on CI to prevent GPU timeout

The diffuse_python_loop benchmark with ns=500 does 4 × 500 Metal
round-trips (laplacian + saxpy) on a 512×512 grid = 2 000 command-buffer
pairs per benchmark run.  The CI paravirtual GPU times out and SIGSEGVs
during this loop (confirmed: crash at cell 28, after cell 26 N-body
benchmark completed fine).

Cap to [10, 50] steps on CI; full [10, 50, 100, 500] runs locally.

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

Author: larsgeb

demo.py

@@ -709,8 +709,10 @@ def diffuse_python_loop(ctx, field, dt, n_steps):
 
 
 # Benchmark: diffuse_steps (C++ loop) vs Python-loop GPU vs NumPy
+# The Python-loop variant does N Python→GPU round-trips per iteration; CI
+# crashes (GPU timeout) at 500 steps, so cap to 50 on the paravirtual device.
 grid_size = 512
-n_steps_list = [10, 50, 100, 500]
+n_steps_list = [10, 50] if _ci else [10, 50, 100, 500]
 dt = 0.1
 
 gpu_fused = []