Bug in symmetry finder

This is a bug in the symmetry finder that we introduced when changing the behaviour of Model. Model no longer removes terms with very small matrix coefficients on initialization. This behaviour was assumed at least at one pont in the code. For example here https://gitlab.kwant-project.org/qt/qsymm/blob/f78723c9bc16fa6383562b1ef58b399f8ee01ab0/qsymm/symmetry_finder.py#L490 we compare the sets of keys on two models, which breaks if new keys appear with small coefficients as a result of cancellation. This particular location is easy to fix, but we should scan the code for errors of this kind.

Below is a minimal example showing this bug. The system has 6-fold rotation symmetry around the z-axis, but it is not found. (Interestingly, the continuous rotation symmetry is found.)

ham = """
    (C + D1*k_z**2 + D2*(k_x**2+k_y**2))*eye(4)
    + kron(sigma_x,sigma_x)*A2*k_y
    + kron(sigma_x,sigma_y)*(-A2*k_x)
    + kron(sigma_y,eye(2))*A1*k_z
    + kron(sigma_z, eye(2))*(M+B1*k_z**2+B2*(k_x**2+k_y**2))
    + m5*kron(sigma_x,sigma_z)
"""
display(qsymm.sympify(ham))

H = qsymm.Model(ham, momenta=['k_x', 'k_y','k_z'])
# candidates = qsymm.groups.cubic()
candidates = qsymm.groups.hexagonal(dim=3)
qsymm.symmetries(H, candidates, continuous_rotations=True, prettify=True)