Fix flaky Lorenz System test with deterministic seed

[ChrisRackauckas-Claude]

Summary

• Add Random.seed!(100) to the Lorenz System test to make neural network weight initialization deterministic
• This prevents intermittent BFGS convergence failures due to unlucky random initialization
• Consistent with other test items in the same file (e.g., direct_function_tests.jl already uses Random.seed!)

Problem

The Lorenz System test in NNPDE2 group was failing intermittently on CI. On the Feb 2 master CI run, all 26 jobs passed except Tests (1.11, NNPDE2) which failed with:

Test Failed at test/additional_loss_tests.jl:185
  Expression: sum(abs2, p_[1] - 10.0) < 100000.0
   Evaluated: 3.1761180413083304e6 < 100000.0

The BFGS optimizer converged to a bad local minimum because the random neural network weight initialization was non-deterministic.

Test plan

• Verified locally: NNPDE2 test group passes (12/12 tests, 0 failures)
• CI should pass on all test groups

🤖 Generated with Claude Code

[ChrisRackauckas-Claude]

CI Results

24/26 test jobs passed. The 2 failures are infrastructure issues, not test code issues:

1. Tests (1.11, AdaptiveLoss): Failed to download GitHub Actions due to API rate limiting (HTTP 429). The lts version of the same test passed fine.
2. Tests (lts, PDEBPINN): The "BPINN PDE Inv III: Improved Parametric Kuromo-Sivashinsky Equation solve" test timed out at 3600s on a slow self-hosted runner. The 1.11 version of the same test passed (59m32s). This test is inherently close to the timeout limit.

All actual test groups pass:

• QA (both versions) ✅
• NNODE, NNPDE1, NNPDE2, NNSDE ✅
• DGM, Forward, IntegroDiff ✅
• AdaptiveLoss, NeuralAdapter ✅
• ODEBPINN, PDEBPINN (1.11), PINOODE ✅
• Integration test, Spell check, Runic format ✅

Please re-run the failed jobs to confirm they're transient.