v4.2.0: Numba JIT compilation for TKE closure

Numba JIT compilation for the Deardorff TKE closure, eliminating the dominant performance bottleneck.

Performance

• Profiling showed the TKE closure consumed 86% of total runtime in both 2D and 3D simulations
• The inner column functions (_compute_mixing_length and compute_tke_closure_column) contained Python-level for k loops over vertical levels
• After JIT compilation: ~200x speedup per column call (from ~1 ms to ~5 μs)

Changes

• Numba @njit on TKE closure column functions — _compute_mixing_length_jit() and _compute_tke_closure_column_jit() accept only NumPy arrays and scalar primitives; Python wrappers preserve the original API for backward compatibility
• config/cbl_3d.yaml — t_end extended from 3600s (1 hour) to 7200s (2 hours)
• requirements.txt — Added numba

Backward Compatibility

• All public APIs unchanged — callers do not need any modifications
• First call incurs a one-time JIT compilation overhead (~1-2s); subsequent calls run at compiled speed
• Numerical results identical to v4.1.0

See RELEASE_NOTES.md for full details.