Give the SRI weak-order study enough trajectories for its own tolerance (#4096)

`iip_weak.jl`'s SRI study asserts `abs(𝒪est[:weak_final] - 2) < 0.5` from 2e4
trajectories. Sampling the estimator over 12 independent seeds shows that
tolerance is too tight for that sample size:

  trajectories   𝒪 range           mean    sd      max |𝒪-2|
  2e4            [1.822, 2.277]    2.006   0.172   0.277
  1e5            [1.892, 2.214]    2.008   0.100   0.214
  4e5            [1.910, 2.078]    1.972   0.049   0.090

The mean sits on 2 at every count and `sd` falls as `1/sqrt(n)` to within noise,
so the estimator is unbiased and the only problem is spread. The study's own RNG
state — seeded once at the top of the file and advanced by the fifteen studies
before it — lands about 3σ out, at `0.5533865655544976`, and fails.

This has not been seen because the group never reached line 186: the file was
OOM-killed at the SROCK2 call on line 71 until #4060.

Raise the count to 4e5 rather than widen the tolerance. At 4e5 the worst of 12
seeds is 0.09 from 2, more than five times inside the existing bound, so the
assertion still catches a real order regression; widening it to admit a bad
estimate would not. The cost is ~4 minutes for that one study.

Verified with the whole group, 16 GiB cgroup, JULIA_NUM_THREADS=2:

  GROUP=IIPWeakConvergence  ->  tests passed, 17m00s, 1.55 GiB peak

which is the first end-to-end run of this file — everything after line 186 had
never executed before.

Fixes #4061



Claude-Session: https://claude.ai/code/session_01UwLXp5WY1uiqPun7qhwxeU

Co-authored-by: ChrisRackauckas-Claude <accounts@chrisrackauckas.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: 3 people

lib/StochasticDiffEq/test/weak_convergence/iip_weak.jl

@@ -179,8 +179,12 @@ sim = test_convergence(
 #@test abs(sim.𝒪est[:weak_l∞]-1) < 0.3
 println("SRI")
 dts = 1 .// 2 .^ (8:-1:2)
+# 2e4 trajectories leaves the weak-order estimate with a standard deviation of 0.17
+# against a 0.5 tolerance, so a ~3σ draw breaches it — which is what this study's own
+# RNG state produced (𝒪 off by 0.553). 4e5 brings the spread to 0.049, worst case 0.09
+# over 12 seeds. See issue #4061.
 sim = test_convergence(
-    dts, prob, SRI(), save_everystep = false, trajectories = Int(2.0e4),
+    dts, prob, SRI(), save_everystep = false, trajectories = Int(4.0e5),
     weak_timeseries_errors = false
 )
 @test abs(sim.𝒪est[:weak_final] - 2) < 0.5