fix: use #|notest instead of #|slow for CI-skipped cells

fasterbench's nbdev settings use tst_flags='notest' (not 'slow'),
so #|slow cells were running in CI and failing on the torchvision
import. Rename to #|notest to match package convention.

Author: nathanhubens

nbs/analysis/roofline.ipynb

@@ -398,42 +398,15 @@
    "id": "test_slow_resnet",
    "metadata": {},
    "outputs": [],
-   "source": [
-    "#| hide\n",
-    "#| slow\n",
-    "from torchvision.models import resnet18\n",
-    "\n",
-    "_model = resnet18()\n",
-    "_sample = torch.randn(1, 3, 64, 64)\n",
-    "_synth = HardwarePeaks(\n",
-    "    peak_flops=1e13, peak_bandwidth=5e11, ridge_point=20.0,\n",
-    "    device=\"cpu\", dtype=\"torch.float32\",\n",
-    "    tf32_enabled=False, cudnn_benchmark=False,\n",
-    ")\n",
-    "_ra = RooflineAnalyzer(_model, _sample, peaks=_synth)\n",
-    "_results = _ra.profile(device=\"cpu\", warmup=2, steps=3)\n",
-    "assert len(_results) > 0\n",
-    "assert all(r.bound in {\"memory\", \"compute\", \"undefined\"} for r in _results)\n",
-    "_ra.summary(top=5)\n",
-    "_fig = _ra.plot()\n",
-    "assert isinstance(_fig, go.Figure)"
-   ]
+   "source": "#| hide\n#| notest\nfrom torchvision.models import resnet18\n\n_model = resnet18()\n_sample = torch.randn(1, 3, 64, 64)\n_synth = HardwarePeaks(\n    peak_flops=1e13, peak_bandwidth=5e11, ridge_point=20.0,\n    device=\"cpu\", dtype=\"torch.float32\",\n    tf32_enabled=False, cudnn_benchmark=False,\n)\n_ra = RooflineAnalyzer(_model, _sample, peaks=_synth)\n_results = _ra.profile(device=\"cpu\", warmup=2, steps=3)\nassert len(_results) > 0\nassert all(r.bound in {\"memory\", \"compute\", \"undefined\"} for r in _results)\n_ra.summary(top=5)\n_fig = _ra.plot()\nassert isinstance(_fig, go.Figure)"
   },
   {
    "cell_type": "code",
    "execution_count": null,
    "id": "test_slow_cuda",
    "metadata": {},
    "outputs": [],
-   "source": [
-    "#| hide\n",
-    "#| slow\n",
-    "if torch.cuda.is_available():\n",
-    "    _p = measure_peaks(device=\"cuda\", matmul_size=512, bandwidth_mb=64, steps=3, warmup=1, cache=False)\n",
-    "    assert _p.peak_flops > 0\n",
-    "    assert _p.peak_bandwidth > 0\n",
-    "    assert _p.ridge_point > 0"
-   ]
+   "source": "#| hide\n#| notest\nif torch.cuda.is_available():\n    _p = measure_peaks(device=\"cuda\", matmul_size=512, bandwidth_mb=64, steps=3, warmup=1, cache=False)\n    assert _p.peak_flops > 0\n    assert _p.peak_bandwidth > 0\n    assert _p.ridge_point > 0"
   },
   {
    "cell_type": "markdown",