Make inference sampling independent of worker scheduling

[Essoz]

Process-based inference previously shared or inherited mutable RNG state. Changing worker count or task scheduling could therefore change sampled evidence and support counts, even for the same saved traces. This gives generation, cross-trace collection, and precondition tasks deterministic RNG streams keyed by their semantic identity. Serial and process paths use the same scopes. Precondition downsampling keeps its existing seed-42 sample choices without reseeding the application's global RNG.

Targets daniel/parallel-infer so this PR contains only the sampling fix and its regression tests. Sampling rates and caps are unchanged; moving to task streams can change the selected subset once compared with the previous serial implementation.

Validation:

• All 27 saved-trace runs agreed on full invariant and failed-hypothesis JSON within each workload: MNIST, Transformer, and DCGAN × 1/2/4 workers × three repetitions. Comparison includes support counts and duplicate multiplicity; only object-key and file-line order are ignored.
• 11 Linux regression tests pass on the final PR files, including actual containment sampling with 1,005 nested calls per trace, fork and spawn, reversed/repeated dispatch, task/thread isolation, frozen exceptions, stable digest seeds, and exact legacy precondition samples above 5,000 examples.
• Commit hooks for Ruff, Mypy, Black, and isort pass, as does git diff --check.
• Old serial versus fixed: predicates and failed hypotheses are unchanged on all three workloads. One positive-support count changes from 20 to 21 in Transformer and one in DCGAN; each is consistent across all patched runs.

The real-workload matrix used Python 3.10.16, PyTorch 2.2.2, pandas 2.3.3, fork, and fixed PYTHONHASHSEED=0. This establishes agreement for the tested inputs and worker counts; it does not claim whole-engine determinism across Python versions, hash seeds, or reordered inputs. Existing set-population ordering and reservoir merge/cap defects remain separate. Two macOS subprocess cases hit OpenMP shared-memory initialization failures; the complete suite passed on Linux.

Reproduce the regression suite:

PYTHONDONTWRITEBYTECODE=1 OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 \
  python -m unittest discover -s tests -p test_inference_sampling.py -v