Fix all open issues: #69, #70, #71

[dantzert]

Summary

Fixes all three open issues and makes the CI green.

Issue #69 — UnderdampedOscillatorKernel overflow for zeta >= 1

• modpods/kernels.py: Replaced the unstable sinh(s*t) formulation in the overdamped branch with a stable difference-of-exponentials form. For zeta > 1 both exponents are negative, so no overflow occurs. The limiting case zeta -> 1 is also handled correctly.

Issue #71 — metrics.py crashes on zero predictions due to log10(0)

• modpods/metrics.py: Refactored compute_detailed_metrics to use a _log10_sorted helper that clips sorted values to a small positive floor (1e-12) before np.log10. The FDC denominator is also guarded against zero. Zero/negative predictions no longer produce -inf/NaN metrics.

Issue #70 — lti_system_gen is extraordinarily slow for realistic problems

• modpods/train.py: Exposed the Bayesian optimization budget (n_calls, n_initial_points) as overrideable optimizer kwargs, so users can trade accuracy for speed.
• modpods/train.py: Added a fast_mode flag to SingleKernelTrainer, MultiKernelTrainer, and delay_io_train that:
  • Uses a cheaper optimizer budget (max_iter // 10, popsize 5)
  • Caps max_transforms at 2 (sufficient for most physical systems)
  • Switches to differential_evolution instead of Bayesian optimization
• modpods/lti.py / modpods/estimator.py: Added fast_mode and optimizer_kwargs parameters to lti_system_gen, LTISystem, and DelayIO for consistency.
• modpods/train.py: Fixed pre-existing mypy errors on differential_evolution calls.

Verification

• All 69 tests pass (including slow tests)
• black, ruff, codespell, and mypy all pass
• fast_mode=True reduces the issue lti_system_gen is extraordinarily slow for realistic problems #70 repro from ~270s to ~2s