Rethink `to_euler_angles`

[moble]

I'm thinking something like this might give us a nicer treatment of beta:

function to_euler_angles_beta_doubled(q::AbstractQuaternion{T}) where {T}
    w, x, y, z = q[1], q[2], q[3], q[4]

    # α and γ from usual formulas (can be outside [0,2π))
    a₁ = atan(z, w)
    a₂ = atan(-x, y)
    α = a₁ + a₂
    γ = a₁ - a₂

    # β magnitude
    c = hypot(w, z)
    s = hypot(x, y)
    β = 2atan(s, c)  # in [0, π]

    # Decide whether to extend β to > π
    # If the "Hopf" product has negative real part, flip β
    if w*x + z*y < 0
        β = 2T(π) - β
    end

    @SVector [α, β, γ]
end

Probably change to have β ∈ (-π, π] or so. The idea here is to put the weird functional behavior at the same point as the weird geometric behavior — when the rotation is exactly a rotation of the z axis onto the -z axis, and the choices of the ways to get there are exactly all the possible unit vectors in the x-y plane. This should minimize the problems with differentiation by putting the bad place far away.