Make zero-seed Sobol sequences reproducible

[AIM-IT4]

Summary

• make a zero Sobol seed deterministic when high-dimensional direction integers must be generated
• preserve the existing sequence for every nonzero seed
• add a regression for the reported Burley 2020 dimension-33 case

Fixes #2732.

Root cause

Jaeckel direction integers are tabulated through dimension 32. Starting at dimension 33, SobolRsg initializes the remaining direction integers with MersenneTwisterUniformRng(seed). That constructor treats 0 as a request for a clock-based random seed, so otherwise identical Burley2020SobolRsg instances diverged only after crossing the tabulated boundary.

For the zero case only, the fallback now uses the Mersenne Twister seed-array constructor with an explicit zero. Nonzero seeds continue through the existing constructor, so their streams are unchanged.

Validation

• before the production change, the regression failed at coordinate 33 (4269245144 != 482177224)
• focused regression: passed (1/1 test case, 2/2 assertions)
• complete LowDiscrepancyTests suite: passed (21/21 test cases, 31/31 assertions)
• complete QuantLib test suite: passed (1,441/1,441 test cases, 13,366/13,366 assertions; 11m16s)
• git diff --check: passed

[coveralls]

coverage: 75.288% (+0.001%) from 75.287% — AIM-IT4:codex/fix-burley-sobol-reproducibility into lballabio:master