Speed up slow tutorial smoke tests (meta-pytorch#3238)

Summary:
Pull Request resolved: meta-pytorch#3238

Improve SMOKE_TEST handling in 8 botorch tutorial notebooks that were
taking >200s in CI. The changes fall into two categories:

1. Added SMOKE_TEST handling where none existed:
   - orthogonal_additive_gp: Added SMOKE_TEST env var check, reduced
     n_train (20→10), n_test (500→50), n_2nd (50→20)
   - max_value_entropy: Added SMOKE_TEST check, reduced num_restarts
     (10→2), raw_samples (512→4), candidate_set (1000→100)

2. Improved existing SMOKE_TEST handling with missed parameters:
   - cost_aware_bayesian_optimization: Reduced budget (25→5) which
     was never gated on SMOKE_TEST
   - multi_fidelity_bo: Reduced batch q (4→2) in both MFKG and EI loops
   - multi_objective_bo: Reduced BATCH_SIZE (4→2), affects all 3
     acquisition functions per iteration
   - saasbo: Reduced final 'predict on test points' section from
     50→10 train/test points
   - closed_loop_botorch_only: Reduced N_TRIALS (3→1 in smoke mode)
   - decoupled_mobo: Reduced COST_BUDGET (54→30)

Measured speedups (test session 13510799038369321):
- orthogonal_additive_gp: 261.8s → 163.6s (1.6x)
- max_value_entropy: 220.2s → 104.6s (2.1x)
- cost_aware_bayesian_optimization: 234.9s → 143.6s (1.6x)
- multi_fidelity_bo: 247.6s → 147.3s (1.7x)
- multi_objective_bo: 232.9s → 134.4s (1.7x)
- saasbo: 231.0s → 126.6s (1.8x)
- closed_loop_botorch_only: 208.8s → 109.7s (1.9x)
- decoupled_mobo: 256.6s → 162.0s (1.6x)

Overall test suite: 580s → 185s (3.1x speedup).
All 64 tests pass.

Reviewed By: esantorella

Differential Revision: D97134436

fbshipit-source-id: cfd6b3f5c295c9dafcb61e3a945e070a536aa0ba

Author: Balandat

tutorials/closed_loop_botorch_only/closed_loop_botorch_only.ipynb

@@ -29,6 +29,7 @@
       "outputs": [],
       "source": [
         "# Install dependencies if we are running in colab\n",
+        "import os\n",
         "import sys\n",
         "if 'google.colab' in sys.modules:\n",
         "    %pip install botorch\n",
@@ -48,6 +49,8 @@
         "torch.manual_seed(0)\n",
         "torch.set_default_dtype(torch.float64)\n",
         "\n",
+        "SMOKE_TEST = os.environ.get(\"SMOKE_TEST\")\n",
+        "\n",
         "%matplotlib inline"
       ]
     },
@@ -98,7 +101,7 @@
         "    )\n",
         "\n",
         "# Generate sparse training data in [0, 1]^5\n",
-        "n_train = 20\n",
+        "n_train = 10 if SMOKE_TEST else 20\n",
         "train_X = torch.rand(n_train, d)\n",
         "train_Y = (test_function(train_X) + NOISE_STD * torch.randn(n_train)).unsqueeze(-1)\n",
         "\n",
@@ -155,7 +158,7 @@
         "\n",
         "# Evaluate both on a dense test set\n",
         "torch.manual_seed(42)\n",
-        "n_test = 500\n",
+        "n_test = 50 if SMOKE_TEST else 500\n",
         "test_X = torch.rand(n_test, d)\n",
         "test_Y = test_function(test_X)\n",
         "\n",
@@ -419,7 +422,7 @@
         "\n",
         "# Generate new training data with the interaction term\n",
         "torch.manual_seed(1)\n",
-        "n_2nd = 50\n",
+        "n_2nd = 20 if SMOKE_TEST else 50\n",
         "train_X_2nd = torch.rand(n_2nd, d)\n",
         "train_Y_2nd = (\n",
         "    test_function_with_interaction(train_X_2nd) + NOISE_STD * torch.randn(n_2nd)\n",
@@ -443,7 +446,8 @@
         "\n",
         "# Compare RMSE on test set\n",
         "torch.manual_seed(42)\n",
-        "test_X_2nd = torch.rand(500, d)\n",
+        "n_test_2nd = 50 if SMOKE_TEST else 500\n",
+        "test_X_2nd = torch.rand(n_test_2nd, d)\n",
         "test_Y_2nd = test_function_with_interaction(test_X_2nd)\n",
         "\n",
         "with torch.no_grad():\n",