Add deterministic kinetic seed extension mode

[Alexander-Mitrofanov]

Adds PredictorSeedExtensionKinetic, available exclusively as --model=X --mode=K. Each seed follows a deterministic greedy path: compare complete energy changes at both ends, accept only strictly downhill moves, and retain the actual visited structures for energy-ranked output and traceback.

--kineticScore=A|B|C selects thermodynamic (default), Manhattan-normalized or asymmetry-normalized ranking. Ties prefer left, smaller total gap, then smaller first-strand gap. Under --outNoLP, a loop-closing pair and its following stack are evaluated and committed together. The predictor respects GU restrictions, separate RNA loop/span limits, regions, output filtering and overlap settings, and validates explicit seed structures. It retains shorter valid prefixes for non-overlapping output and prevents rejected explicit seeds from returning as seed annotations.

The scientific scope and detailed implementation decisions are recorded in the design and implementation plan. This is a zippering-inspired energy heuristic, without calibrated transition rates, a time axis or a global-optimum guarantee. Equilibrium partition/probability requests are rejected rather than inferred from selected paths; ordinary energy trackers remain supported.

Loop-only energetic pruning is deliberately omitted: a Turner2004 loop costing +0.50 kcal/mol followed by a −3.30 kcal/mol stack gives a favorable complete move that the proposed bare-loop bound would reject. Terminal and both weighted dangling contributions also belong to the exact change. Feasible moves are evaluated exhaustively with the active energy model; no ED monotonicity or default Turner parameters are assumed. No performance improvement over existing predictors is claimed.

Validation:

• GCC 14.4 release and debug builds, ViennaRNA 2.7.2, Boost 1.85, Kokkos mdspan backend.
• Full make tests passed in both builds: 74,114 assertions in 59 API cases, and 5/5 test-suite entries (API plus all CLI scripts).
• Independent absolute-endpoint/whole-chain oracle; exact Turner loop/stack rescue; strict stopping; A/B/C and tie order; macro denominators; asymmetric bounds and nonmonotone ED; full dangle-aware energy; transient GU recovery; invalid seeds; retained paths, annotations and non-overlapping prefixes; repeated calls and zero/boundary-only output.
• CLI energy/coordinate expectations and independent reevaluation of predicted structures through --rri, including model/score/output validation.
• Installation, standalone compilation of the installed new public header, and a compiled/linked/running installed-library consumer (with the documented Easylogging initialization and library).
• git diff --check passed. Existing regression reference outputs were unchanged.

README, CLI help, distribution registration and ChangeLog are included. The full-path cache has memory proportional to the sum of retained path lengths; the design documents this tradeoff.