ResolutionRenderer only works for square observations

[pmelchior]

Follow up from #276

The immediately offending line is

scarlet/scarlet/renderer.py

Lines 274 to 275 in 5b0b5e0

	pixels = np.stack((np.arange(lr_shape[0]), np.arange(lr_shape[1])), axis=1)
	coord_hr = data_frame.convert_pixel_to(model_frame, pixel=pixels)

What's going on here is the Y and X coordinates of the observation are stacked together and then transformed to model coordinates. This should probably be better done with two different arrays, one with the 2D coordinates of the Y-axis and one the X-axis, instead of using the diagonal, which is what triggers the error.

But doing that doesn't solve the problem when the frames are rotated. To go from the rotated to the unrotated coordinates is done here:

scarlet/scarlet/renderer.py

Lines 322 to 329 in 5b0b5e0

	Y_unrot = (
	(coord_hr[:, 0] - center_y) * self.angle[0]
	- (coord_hr[:, 1] - center_x) * self.angle[1]
	).reshape(lr_shape[0])
	X_unrot = (
	(coord_hr[:, 1] - center_x) * self.angle[0]
	+ (coord_hr[:, 0] - center_y) * self.angle[1]
	).reshape(lr_shape[1])

This requires that both coordinates in coord_hr have the same length, i.e. that their frames are square. That's consistent with eq 20 in Remy's paper, where X and Y coordinates also have to be of same length.

This issue is here to document to problem. We could allow for non-square frame in the non-rotated case, but we must assert that the observations are square in the rotated case.