Seed every side of the benchmark and Table 1 protocols

[neuromechanist]

Found in epic #324 Phase 15 (#351).

Some sides of these protocols are not seeded or not matched:

• benchmarks/reproduce_table1.py: the reference runs in the single-model tier and the multi-model ensemble draw clock-based seeds. On the bundled tier, one reference run of five landed in a lower-likelihood basin, which a rerun may or may not reproduce.
• The ensemble passes invsigmin 0.0 to the reference (as runamica15.m writes it with %f) and 1e-8 to pamica.
• benchmarks/benchmark_dimsweep.py calls AMICA_NumPy without a seed, so its log-likelihood column starts from a different draw than the other backends.

Proposal:

• Seed the reference side (the seed keyword) and the NumPy side.
• Align invsigmin.
• Re-run .context/issue-351/pin_check.py --dump whenever AMICANative's parameter table changes.
• Changing the Table 1 protocol means re-measuring its rows; do it together with the next re-measurement.