Use scattering angle function from ScippNeutron

[jl-wynen]

Similarly to scipp/esssans#143, use some code in scippneutron.conversion.beamline to compute theta. We can't just use scattering_angles_with_gravity because in reflectometry, we define the scattering angle w.r.t. the sample plane, not the beam direction. Compared to SANS, this means that we ignore the x component in y_prime for two_theta.

[jl-wynen]

I believe I found a bug in the current implementation. We currently define
$$y_d = \frac{g \cdot b_2}{|g|}$$
where $g$ is the gravity vector (pointing down) and $b_2$ is the scattered beam.
see

essreflectometry/src/ess/reflectometry/conversions.py

Line 55 in bddec78

y = sc.dot(scattered_beam, gravity) / grav

This definition means that the y axis points down. I.e., positive y values are below the incoming beam and negative values above the incoming beam. We then compute the point where the neutron would have been detected without gravity as
$$y'_d = y_d + \frac{|g| m_n^2}{2 h^2} L_2^2 \lambda^2$$
This equation is identical to that in SANS (scattering_angles_with_gravity). However, in SANS, we define (note the minus sign!)
$$y_d = -\frac{g \cdot b_2}{|g|}$$
see https://github.com/scipp/scippneutron/blob/1a1c86a65fb69c727f4c28bef05d8a063e66bb2f/src/scippneutron/conversion/beamline.py#L384

The definition in SANS makes sense. It means that the direction the neutron was scattered in ($y'_d$) is above where it was detected ($y_d$) where 'above' means in negative $g$-direction. But in reflectometry, we get the opposite. The gravity corrected value is below the detected position, i.e., further down along $g$. This is wrong.

This may not have been noticed before because we use an absolute value of $y'_d$ which might mask the effect. But I still think that the result is wrong. The impact on the result is significant:

Has this been compared to Mantid? @jokasimr, @arm61 do you agree with this?

[jl-wynen]

Here is a comparison of IofQ:

[jokasimr]

Yes it looks like this could be an issue.

The results have not been compared to mantid. But they have been compared to Jochens program, and there we see a clear discrepance in the form of a shift to lower $Q$. It seems the discrepancy increases with the sample rotation angle is large. See https://scipp.github.io/essreflectometry/user-guide/amor/compare-to-eos.html.

However, the (possible) gravity correction bug here is not enough to explain the discrepancy.