demo: remove 50M-element benchmark size that OOMs the CI paravirtual GPU

The 1D benchmark ran up to N=50_000_000 (200 MB per array, 600 MB for the
three Metal buffers simultaneously in flight).  The CI runner's paravirtual
GPU cannot allocate that much Metal shared memory and crashes with SIGSEGV.

N=10_000_000 (40 MB per buffer) already captures the crossover point where
GPU wins over NumPy and keeps total memory well within the runner's limits.

Also remove the per-cell progress logging that was added for diagnosis.

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

Author: larsgeb

demo.py

@@ -117,7 +117,6 @@ def benchmark(func, *args, repeats=100, warmup=15):
     100_000,
     1_000_000,
     10_000_000,
-    50_000_000,
 ]
 ops = {
     "add": {

scripts/render_notebooks.py

@@ -82,10 +82,6 @@ def _patched_show(*_args, **_kwargs) -> None:
         if not src:
             continue
 
-        # Progress line bypasses stdout_buf so it survives a crash mid-cell.
-        preview = src.splitlines()[0][:60]
-        print(f"  [cell {exec_count}] {preview}", file=sys.__stdout__, flush=True)
-
         cell_images.clear()
         stdout_buf = io.StringIO()