Speed up test_weakref_leak and test_step_args

Profiling-backed changes targeting the two tests where iteration
reduction yields meaningful savings.

test_weakref_leak (89s on CI → estimated ~30s):
Object count profiling shows memory state stabilizes at iteration 2.
The original test used 16 warmup iterations before checking — reduced
to 3 warmup + 3 check = 6 total iterations (from 20). Each
conditional_logp call costs ~4.5s on CI, so removing 14 iterations
saves ~63s.

test_step_args (62s on CI → estimated ~25s):
This test verifies target_accept argument plumbing, checking
acceptance_rate.mean() ≈ 0.5 with decimal=1 precision. The default
pm.sample() uses 1000 draws/tune, but 200 is sufficient for this
loose tolerance. Verified stable over 10 runs with different seeds.

Changes NOT made (profiling showed negligible impact):
- test_default_value_transform_logprob range(10)→range(3): compile
  time is 99.98% of cost, loop saves 0.6ms total
- test_interpolated x_points 100k→20k: compilation is 91.8% of cost,
  array reduction saves ~296ms across 56 iterations

Addresses #7686

Author: mcwanza

tests/logprob/test_transform_value.py

@@ -634,9 +634,8 @@ def _growth(limit=10, peak_stats={}):
     rvs_to_values = {pt.random.beta(1, 1, name=f"p_{i}"): pt.scalar(f"p_{i}") for i in range(30)}
     tr = TransformValuesRewrite(dict.fromkeys(rvs_to_values.values(), logodds))
 
-    for i in range(20):
+    for i in range(6):
         conditional_logp(rvs_to_values, extra_rewrites=tr)
         res = _growth()
-        # Only start checking after warmup
-        if i > 15:
+        if i > 2:
             assert not res, "Object counts are still growing"

tests/sampling/test_mcmc.py

@@ -691,10 +691,12 @@ def accept(idata):
 
     with pm.Model() as model:
         a = pm.Normal("a")
-        idata_default = pm.sample(random_seed=1410)
-        idata0 = pm.sample(target_accept=0.5, random_seed=1410)
-        idata1 = pm.sample(nuts={"target_accept": 0.5}, random_seed=1410 * 2)
-        idata2 = pm.sample(target_accept=0.5, nuts={"max_treedepth": 10}, random_seed=1410)
+        idata_default = pm.sample(draws=200, tune=200, random_seed=1410)
+        idata0 = pm.sample(draws=200, tune=200, target_accept=0.5, random_seed=1410)
+        idata1 = pm.sample(draws=200, tune=200, nuts={"target_accept": 0.5}, random_seed=1410 * 2)
+        idata2 = pm.sample(
+            draws=200, tune=200, target_accept=0.5, nuts={"max_treedepth": 10}, random_seed=1410
+        )
 
         with pytest.raises(ValueError, match="`target_accept` was defined twice."):
             pm.sample(target_accept=0.5, nuts={"target_accept": 0.95}, random_seed=1410)
@@ -709,13 +711,17 @@ def accept(idata):
     with pm.Model() as model:
         a = pm.Normal("a")
         b = pm.Poisson("b", 1)
-        idata0 = pm.sample(target_accept=0.5, random_seed=1418)
+        idata0 = pm.sample(draws=200, tune=200, target_accept=0.5, random_seed=1418)
         with warnings.catch_warnings():
             warnings.filterwarnings(
                 "ignore", "invalid value encountered in double_scalars", RuntimeWarning
             )
             idata1 = pm.sample(
-                nuts={"target_accept": 0.5}, metropolis={"scaling": 0}, random_seed=1418 * 2
+                draws=200,
+                tune=200,
+                nuts={"target_accept": 0.5},
+                metropolis={"scaling": 0},
+                random_seed=1418 * 2,
             )
 
     npt.assert_almost_equal(accept(idata0).mean(), 0.5, decimal=1)