Simplify and generalize Helmert logic

[danvk]

Follow-up to #385 and #388. See also #381.

This streamlines the Helmert calculations so that they all go through numpy's np.linalg.lstsq. Since Helmert is a fundamentally linear transformation, this is valid and will produce exact fits for two GCPs and least squares fits for more. No special-casing or if statements needed.

This also fixes a bug where _calculate_scale only considered the first two GCPs but the other methods considered all of them. I made _get_helmert_params return a dataclass with the parameters. The test passes without much modification, just an assertEqual -> assertAlmostEqual for the rotation.

[mradamcox]

@danvk Thanks for this, it's definitely a more concise approach to generating these 4 params, and after testing it all out manually (I wasn't fully confident my unit tests would be sufficient) everything still works as expected. I hadn't really considered that scale would need to take into account more than two GCPs if they were present, but you're right that it does.

Can you go ahead and also remove the geometry.azimuth_from_coords() function? There is no need for it anymore, though I did enjoy figuring it out :).

[danvk]

@danvk Thanks for this, it's definitely a more concise approach to generating these 4 params, and after testing it all out manually (I wasn't fully confident my unit tests would be sufficient) everything still works as expected. I hadn't really considered that scale would need to take into account more than two GCPs if they were present, but you're right that it does.

Glad to hear it's working as expected! This is definitely the right way to fit a model.

Can you go ahead and also remove the geometry.azimuth_from_coords() function? There is no need for it anymore, though I did enjoy figuring it out :).

Done, now we've passed -100 ∆ lines!

[mradamcox]

Great, thanks!

Unfortunately I've found there is still more work to do to bring the helmert transformation into the IIIF annotation output, as described here: #391