Imaginary vibrational modes silently produce NaN thermochemistry

[galjos]

An imaginary (negative-wavenumber) mode in the kept dof set makes the harmonic-oscillator formulas in thermo.py (_compute_vibrational_*, ~lines 293-398) return nan for entropy (np.log(1 - exp(-theta/T)) of a negative argument) and a downward-corrupted ZPE/internal energy. System.has_imaginary_frequencies is computed but Thermo never consults it, so the result is a silent nan total entropy / Gibbs rather than an error.

Repro:

import numpy as np
from ThermoScreening.thermo.atoms import Atom
from ThermoScreening.thermo.system import System
from ThermoScreening.thermo.thermo import Thermo

atoms = [Atom(symbol=s, position=np.array(p, float)) for s, p in
         zip(["O","H","H"], [[0,0,0.119],[0,0.763,-0.477],[0,-0.763,-0.477]])]
system = System(atoms, periodicity=False, cell=None, charge=0,
                electronic_energy=0.0, vibrational_frequencies=np.array([-500.0, 3657.0, 3756.0]))
t = Thermo(temperature=298.15, pressure=101325, system=system, engine="dftb+"); t.run()
print(system.has_imaginary_frequencies, t.total_entropy("cal/(mol*K)"))
# True nan

Imaginary modes are common in real DFTB+ output (transition states, unconverged minima), so silent NaN is a real hazard for screening.

Suggested fix: have _vibrational_contribution consult system.has_imaginary_frequencies and raise a TSValueError ("Imaginary vibrational frequencies are present.") when an imaginary mode is in the kept dof set (or explicitly exclude/floor them, with a documented policy). The HO formulas themselves are correct. Found alongside #59 (rotational geometry fix).