Fix premature truncation in double no-touch pricing

[kadyrbekovhamit-cyber]

Closes #266.

A nonnegative cash payoff can receive a negative price: the reported 30-day example returns no-touch 1.1717090895 and touch -0.1733515747, although discounted cash is 0.9983575147.

The series can stop at an individual zero coefficient or sine term while later terms remain material. Its damping-based cutoff also computes half the threshold implied by the documented exponential bound. This change removes the single-term stop and corrects that factor, retaining the existing 50-term minimum and 2,000-term cap. No output clipping is introduced.

The new deterministic tests cover pricing bounds, touch/no-touch parity, an independent reflected-Gaussian reference, short and long maturities, payment scaling, and nonzero log drift.

Validation on current master 4c7cd50bdadd6efc9ac74fa81e93374397d18e3e:

• All 89 new cases pass; the original fails 56, removing only the early stop still fails 13, and restoring the original reproduces the same 56 failures. Released 1.1.2 also fails the same cases, with a test-only import alias for the renamed flat curve.
• Exact PR files plus 14 existing adjacent FX tests: 103 passed.
• A separate 480-scenario public-API matrix, with a freshly recomputed 60-decimal method-of-images oracle, reproduces the retained results: 360 failures originally, 48 after the first change, zero after both, and 360 after restoration.

The 2,000-term cap is unchanged; this is not a universal convergence guarantee. The full upstream suite, Monte Carlo, Greeks and performance were not measured. Implementation and test preparation were AI-assisted; validation used the actual Numba CPU implementation.

[domokane]

Thanks. I fixed this manually before I saw this detail.