Tighten SFS migration integration accuracy

[terhorst]

Summary

• tighten the Diffrax tolerances used for continuous-migration SFS lifts
• let the adaptive controller choose its initial step so accuracy does not depend on epoch partitioning
• make the equal-rate one-epoch/two-epoch test robust in both float32 and float64

Investigation

• demestats and Moments migration matrices agree to machine precision
• Moments pure-migration results converge to demestats as dt_fac is reduced
• pulse-only demestats results pass the existing momi2 parity tests
• the American-admixture discrepancy is caused by Moments globally upsampling an independent low-sample deme; matched-grid results converge from 9.2e-5 TV at grid 4 to 2.2e-5 at grid 8

Tests

• focused migration tests: 4 passed in float32
• focused migration tests: 4 passed in float64
• pulse/momi2 tests: 10 passed, 5 skipped