Speed up test_weakref_leak and test_step_args

[mcwanza]

Addresses #7686 — profiling-backed optimizations targeting tests where iteration reduction yields meaningful savings.

Changes

test_weakref_leak (89s → ~30s estimated on CI)

Object count profiling shows memory state stabilizes at iteration 2:

Iteration | Growing objects | Details
    0     | 10 growing      | function:+67554, tuple:+50400, dict:+30266
    1     |  1 growing      | list:+1
    2     |  0 growing      | STABLE
    3     |  0 growing      | STABLE
   ...    |  0 growing      | STABLE (all remaining iterations)

The original test ran 20 iterations but only checked after iteration 15 — 16 warmup iterations when 3 is sufficient. Reduced to 6 total (3 warmup + 3 check). Each conditional_logp call costs ~4.5s on CI, so removing 14 iterations saves ~63s.

test_step_args (62s → ~25s estimated on CI)

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 more than sufficient for this loose tolerance (verified stable over 10 runs with different seeds).

Profiling: first pm.sample() with 1000 draws takes 2.20s locally, with 200 draws takes 0.29s. The test calls pm.sample() 5 times.

Changes NOT made (profiling showed negligible impact)

Test	CI time	Bottleneck	Iteration savings
test_default_value_transform_logprob	35s	Compile: 8.85s, Loop(10): 0.0014s	0.6ms total
TestMatchesScipy::test_interpolated	33s	Compile: 91.8%, Array: 7.9%	~296ms across 56 iters

Compilation dominates these tests — reducing iteration counts has near-zero impact, consistent with @ricardoV94's review feedback.

[ricardoV94]

Some changes are fine, other sacrifice too much on the tests. Left some commenst

[ricardoV94]

I don't think we should remove any of the parametrizations in test_multivariate. If these are slow the best thing is to analyze why instead

[mcwanza]

Reverted — all test_multivariate.py changes have been removed from this PR.

[mcwanza]

Reworked the entire PR based on this feedback. Profiled all the slow tests to understand where time is actually spent — you were right that compilation dominates.

Removed all the original changes (parametrization reduction, iteration count tweaks on compile-dominated tests). The PR now only contains 2 profiling-backed changes:

1. test_weakref_leak (20→6 iterations): Object count profiling shows state stabilizes at iteration 2 — the original 16 warmup iterations were overkill. Each conditional_logp call costs ~4.5s on CI, so this saves ~63s.

2. test_step_args (explicit draws=200, tune=200): This test checks acceptance_rate.mean() ≈ 0.5 with decimal=1 precision — 200 draws is sufficient. The default 1000 was unnecessary for this tolerance level.

[ricardoV94]

I would be very surprised if this played a role. The cost will be mostly compile time, not running 5 +- steps. Let me know if I am wrong

[mcwanza]

Reverted — the progress bar change has been removed from this PR.

[mcwanza]

Confirmed by profiling. The test_default_value_transform_logprob loop takes 0.0014s total for 10 iterations while compilation takes 8.85s — compile/loop ratio of 6480x. Removed this change from the PR.